Methods

Study population

The SMP measurements were carried out within the framework of INTERPHONE, an international case-control study investigating whether RF exposure from mobile phones is related to risk of tumours of the brain, acoustic nerve and parotid gland.37 SMPs were used by volunteer subjects in 12 of 13 INTERPHONE countries: Australia, Canada, Denmark, Finland, France, Germany, Israel, Italy, New Zealand, Norway, Sweden and the UK, between 2001 and 2005. Volunteers were generally colleagues and acquaintances of the investigators. It was usually not feasible to recruit random samples of mobile phone users from the general population because of the valuable equipment (SMPs) used. In each study centre, an effort was made to select volunteers to correspond broadly to the INTERPHONE study sample with respect to age and gender. In Sweden, volunteers were also selected to cover both urban and rural mobile phone users. The other centres covered mostly urban and suburban areas to reflect the main INTERPHONE study sample.37 The protocol called for approximately 50 volunteers to be recruited in each centre. People who only used the SMPs for a few days were excluded (n = 25). The SMP studies were approved by the IARC Ethical Review Committee and by the relevant ethical committees of the participating centres.

Data recorded by the SMPs

The SMPs had the appearance of normal mobile phones, but contained software to internally record data. Four SMP models (25–35 handsets of each model) were developed by four major mobile phone manufacturers for this study. All phones were dual (900/1800 MHz) or tri-band (900/1800/1900 MHz). Volunteers put the subscriber identity module (SIM) card from their usual mobile phone in the SMP and continued making calls using their own phone number and subscription. Usually, two models of phones were distributed in each study centre. Each volunteer used the SMP for around 1 month.

All four phone models recorded the date, time and duration of each call made or received, and the frequency band (900 or 1800/1900 MHz) and power level in each band, at intervals throughout the call. The GSM standards define 15 power levels in the 900 MHz band (from 33 to 5 dBm or 2 to 0.003 W) and 16 levels in the 1800 MHz band (from 30 to 0 dBm or 1 to 0.001 W).31 Two of the phone models had sampling intervals of 2.5 s; the third sampled every 120 ms. The fourth phone model recorded every time there was a change in one of the recorded variables (band and power level), with a minimum sampling interval of 1 s. Analyses included data recorded during speech communication only. The SMPs did not record information about DTX (discontinuous transmission mode).

Questionnaire data

Information about the main circumstances of phone use of interest for exposure assessment was collected by questionnaire. Participants were asked about the frequency of phone use in urban or rural areas and in a moving vehicle, using the same questions as the main INTERPHONE study questionnaire.37 An additional question about use of the phone inside buildings was later added to the SMP questionnaire after a re-evaluation of its possible importance (not used in five centres: Denmark, France, Germany, Sweden, UK North).

The core protocol for the SMP studies called for this questionnaire to be administered at least 6 months after the use of the SMPs, as one of its other aims was the testing of short-term recall.23 Some centres (Canada, Denmark, France, Germany and the UK) also administered short questionnaires immediately after the SMP use period. In these countries the questionnaire nearest to the time of SMP use was used.

Calculation of average power levels

The power level recorded for each sampling period in a call was converted to output power in milliwatts (mW) and divided by a factor of 8 to allow for time division multiple access (TDMA)27; maximum output powers were thus 250 mW at 900 MHz and 125 mW at 1800 MHz. Average output powers for a call were calculated separately for each frequency band within a call. Phone calls using both frequency bands thus had two average output powers, one for each band. In addition to average output power per call, we calculated for each phone call the percentage of call time per call during which the phone emitted at maximum power (PMax) and at minimum power (PMin), combining data from the two frequency bands. Phone calls made entirely at PMax were identified.

Analyses

Main analyses used linear regression models to evaluate the influence of a set of explanatory variables on the average output power per call and the percentage of call time at PMax. The following explanatory variables were included (see table 1): study centre, network operator, time period of study, SMP model, call duration, time of day, use in urban or rural areas, use in a moving vehicle, use inside buildings, gender and age. Call duration and time of day were available for each phone call; all other variables were available at the aggregate level of the individual. The measures of effect presented are the regression coefficients, which reflect the absolute difference in average output power (in mW) or in percentage of call time at PMax between different categories of the explanatory variables. Calls were used as the unit of observation in analyses throughout. All regression analyses used a random effects model to account for the multiple calls made by one subject (Stata command xtreg38). Each explanatory variable was evaluated separately in a model adjusting for operator and call duration, as these were the most important predictors of the output power, and for SMP model, because power levels may be related to the design of a handset.

The outcome variables (average power levels and PMax) were not normally distributed; rather, they followed bimodal distributions with peaks at the minimum and maximum levels. In these circumstances it is important to test the sensitivity of the linear regression models to normality assumptions. For these sensitivity analyses, we used ordinal logistic regression (or proportional odds) models (STATA command ologit). We categorised each outcome variable into six ordinal categories: 0–24, 25–74, 75–124, 125–174, 175–224, 225–250 mW for the average power level at 900 MHz; 0–12, 13–37, 38–62, 63–87, 88–112, 113–125 mW for the average power level at 1800 MHz; and 0–9, 10–29, 30–49, 50–69, 70–89, 90–100% for PMax. The explanatory variables were the same as in the linear models.

We further examined the contribution of each explanatory variable to the variance of a model containing all variables. In clustered models, the calculation of goodness-of-fit or R² is not as straightforward as in standard linear regression models; Bayesian models have been proposed to calculate R² values for different levels of a multi-level model,39 but their application falls outside the scope of the current paper. We therefore chose to report the percentage change in the R² within and between subjects (Stata command XTREG38), removing one variable at a time from the full model. These are unadjusted R² and not to be interpreted as the absolute fraction of explained overall variance as in non-clustered models, but they give an indication of the relative contribution of each variable to, respectively, the intra- and inter-subject variance.

Results

Analyses were based on a total of 516 subjects from 12 countries who made or received 63 151 calls with the SMPs (table 2). The average length of SMP use was around 30 days. Overall, 47% of users were male. Sixty-six per cent were aged between 31 and 50 years, and 15% were aged <30 years old, an age group not included in the main INTERPHONE study. The average number of calls made per day varied considerably by country, from 1.6 in the first UK study to 10.9 in Israel. The average duration of calls varied much less and ranged from 1.5 minutes in Germany to 2.8 minutes in Finland (table 2).

Study centre and network

For all study centres combined, the mean output power per call was 133.3 mW at 900 MHz and 64.2 mW at 1800 MHz (table 3), each approximately 50% of the maximum power levels of 250 and 125 mW, respectively. The mean output power ranged from 92 mW in Israel to 171 mW in New Zealand at 900 MHz and from 47 mW in Israel to 88 mW in Finland at 1800 MHz. Study centre differences were highly statistically significant (likelihood ratio; LR test, p<0.001) even when centres with the most extreme mean values (eg, Israel) were excluded. There was wide variability in output power, with standard deviations between 50 and 80% of the mean. The phones emitted at PMax on average for 39% of call time and at PMin for 2.8% (table 3 and Appendix fig 1A), again with substantial differences between study centres (table 3 and Appendix fig 1B). In all centres combined, 21% of phone calls were made entirely at PMax (table 3).

Overall, the mean output powers varied by a factor of 2 (at 1800 MHz) to 3 (at 900 MHz) between operators (interoperator range shown in table 3). In all regression models network operator was a significant explanatory variable (LR test, p<0.001). Within some, but not all study centres, large variations were observed between operators, but these differences were generally not consistent between the two frequency bands (table 3). Mobile phone output powers did not vary substantially by time period of SMP use, indicating little difference between earlier (2000/2001) and later (2004/2005) years.

Call characteristics

Call duration was a strong predictor of the average output power of a call (table 4). Output power was higher for short than long calls, with an average difference of 27.8 mW (at 900 MHz) and 14.3 mW (at 1800 MHz) comparing calls of less than 5 s with calls of more than 60 s. Use of PMax was 16.5% higher in the shortest compared with the longest calls. The effect of call duration on output power appeared to level off after 10–30 s.

The time of day at which the call was made had a much smaller influence on the output power: calls made during the day had a significantly lower output power at 900 MHz, and a lower percentage at PMax, but the absolute differences were very small (table 4).

Phone use circumstances

In all study centres together, calls made by SMP users who reported using their phones mainly in rural areas had somewhat higher mean output power and use of PMax than those reporting use in urban areas (table 4). These differences were statistically significant in the 900 MHz band (p = 0.011) and for PMax (p = 0.007), but they were small compared with the standard deviations around the mean power. Among mainly urban users there were no differences in output power levels between those who replied “city centre” as their main location of phone use and those who replied “suburban” (not shown). Since the concept of what constitutes an urban versus rural area is likely to vary between regions and countries, the urban-rural comparisons were also analysed for each centre separately. Results were somewhat inconsistent between centres (heterogeneity p value: 0.13 at 900 MHz, 0.09 at 1800 MHz) and based on very few rural users in most centres (not shown). Sweden was the only study centre where statistically significantly higher output powers were observed for rural users in both frequency bands (+33.7 mW at 900 MHz (p = 0.003), +24.1 mW at 1800 mW (p = 0.008), +12.7% for PMax (p = 0.005), based on 12 rural users). The Swedish results were also found to drive the results of the overall comparison of urban and rural users: when Sweden was excluded from these analyses, we observed smaller and non-significant differences between urban and rural categories (regression coefficient rural vs urban: 10.3, 95% CI –5.7 to 26.3 at 900 MHz; 1.3, 95% CI –9.4 to 12.1 at 1800 MHz; and 6.1, 95% CI –1.0 to 13.3 for PMax).

Mean mobile phone output power appeared not to be influenced by reported use while moving in a vehicle or inside buildings; differences in average output power and percentage of PMax between categories of these variables were small and had wide confidence intervals (table 4).

SMP model and user characteristics

Mobile phone output power did not show significant or substantial variations by age, gender, or SMP model (not shown).

Relative importance of each explanatory variable

The relative contribution of each of the explanatory variables to the between- and within-subject variance of a model containing all our explanatory variables confirms the above results (table 5). Of the variables that can only vary between (not within) subjects, operator was the most important contributor to variance between subjects, followed by study centre. Circumstances of phone use had little influence, as seen above. Only duration and time of call varied within subjects (between calls) and duration of call was the most important contributor to variance within subjects.

Sensitivity analyses

Regression models using the ordered, categorised outcome variables gave very similar results to the linear regression models, suggesting that results from the linear models were valid. Explanatory factors with statistically significant effects in the linear regression models (operator, call duration, urban-rural location) had statistically significant effects in the ordinal models. For example, in the ordinal logistic regression rural users were more likely than urban users to use higher power levels and this effect was significant for all three outcomes (p<0.001 for 1900 MHz and PMax, p = 0.01 for 1800 MHz).

Discussion

We measured the output power of more than 60 000 phone calls made by over 500 people worldwide. APC employed by the second-generation GSM networks reduced the output power of mobile phone handsets by a factor of 2 on average. The main predictors of output power were the study location, the network and the duration of the call, rather than circumstances of use as reported by an individual. Reported use of the mobile phone in rural areas was related to higher output power than use in urban areas, principally in the one centre that covered very sparsely populated areas.

APC and circumstances of use

Average output powers in this study were around 50% (ranging from 35 to 70% depending on the country) of the maximum power level. Maximum power levels were used for a large percentage of call time, and a considerable proportion of calls, including longer duration calls, showed no downregulation of output power at all. This study covered limited time periods and geographical areas, but the findings are fairly consistent, with at most a factor of two differences in output power between study centres and operators. The findings therefore suggest that the reductions in power level actually achieved by second-generation GSM networks are much less than the orders of magnitude that are theoretically possible.31

The few prior studies of mobile phone output power levels are generally of more limited scope than the current study. Recordings from a test mobile system in Paris in 1999 concluded that mean power levels were far below the maximum (around 25% of the maximum),27 whereas a more recent US SMP study of over 2500 calls recorded average output powers of around 50% of the maximum.36 A study in Sweden recorded all calls from one operator for 1 week in specific areas and reported frequent use of the maximum power level in rural areas (around 50% of call time), but far less in urban areas (25%).32 A study using call-logging to obtain information on a relatively small number of phone calls in Sweden and the UK came to similar conclusions.35 A small SMP study in Italy found maximum power levels to be used for a large proportion (>40%) of the time, regardless of the location of the call.33 Taken together, these results are consistent with the current study in showing that APC reductions in power levels are likely to be around two- to four-fold.

There are several possible explanations for the relatively modest power reduction achieved by APC. Networks are now configured to use many handovers between base stations within one phone call, even when a user is stationary, to optimise its base station utilisation33; every such new connection is made at high power. Other explanations could be the increased use of phones inside buildings, the shielding of signals by buildings in urban areas, and the very dense traffic of calls leading to base stations reaching maximum capacity with calls being connected to stations further away.33 36

We tested the hypotheses relating higher power levels to the frequency of phone use in rural areas, in moving vehicles, and inside buildings. Overall, these factors showed little influence on output powers. Only in Sweden, where the selection of volunteers covered sparsely populated rural areas, was the output power for rural users higher (by 30–50% depending on band) than for urban users. Other centres covered mostly urban and suburban populations to reflect the main INTERPHONE study population in those centres.37 Our findings for Sweden are consistent with other studies in Sweden.32 35 Studies in the USA and Germany have not reported higher output power in rural situations.34 36 Our lack of a detectable influence of use while moving and use inside buildings is consistent with studies in Sweden, the USA and Germany.34 35 36 The small Italian study observed substantially (68%) higher power levels indoors than outdoors, and in calls made while stationary compared with moving.33

That information on phone use circumstances was available from questionnaire answers only and not for each individual phone call, is a limitation of the current study. Questionnaires about usual use patterns provide only a crude measure of the circumstances of each phone call and may have obscured the real effects of these circumstances on the output power. On the other hand, this design has allowed us to study much larger numbers of phone calls and time periods than previous studies, and, most importantly, has provided us with information that is directly relevant to questionnaire responses obtained from cases and controls in epidemiological studies, because information on single calls cannot be obtained when assessing an individual’s lifetime history of mobile phone use.

Network and study location

Although power reductions achieved by APC were relatively modest, geographical area (study centre) and the network operator used were strong determinants of the output power. The effects of study location and network characteristics are hard to disentangle. Some centres included only one operator, making the two inseparable (non-identifiable). Within some other centres, substantial differences between operators were found, indicating that differences in configuration and operation of the networks may have been responsible. Factors unrelated to the network used cannot, however, be excluded as explanation for the differences found between centres. It should further be noted that this SMP study covered some but not all study regions of INTERPHONE37 and never entire countries, and since network configurations are likely to vary between geographical areas within the same country, inferences about operators or countries cannot be generalised widely.

Call characteristics

As expected, call duration was a determinant of output power. Very short calls were made at higher average power because the first seconds of a call are used to establish the connection at maximum power. The time of day at which a phone call is made might be an indicator of traffic density at the time of the call and thus of the frequency of handovers between base stations; calls during daytime would be expected to use higher power levels, but we did not find higher power levels during daytime. The Swedish base station studies similarly found little difference in power level distributions between day- and night-time calls.32 35

Phone characteristics

The design of a mobile phone handset could influence its output power level and consequently the results of this study. The telephone communications power (TCP) quantifies the power left for communication purposes after a proportion of the radiated power has been absorbed in the head of the user.40 The TCP varies from phone to phone depending on the design. Phones with a higher TCP would be able to operate at lower nominal power levels than other phones while achieving satisfactory signal strength at the receiving base station. Also, GSM standards allow up to SD 2.5 dB variation from the nominal values (33 dBm at 900 MHz and 30 dBm at 1800 MHz) in the actual maximum output power of individual phones.27 We measured the TCPs of one handset of each phone model in the study and found these to be within the range of typical values quoted for phones. SAR values (measuring the energy from the phone absorbed in the head) were also not unusually high or low. These measurements, together with results from our regression models showing no substantial differences in output power between the four phone models, suggest little influence of factors related to handset design on the results of the study. These conclusions only apply for the four models used in this study, however. We had too few models to allow for a general evaluation of the influence of phone model on power output levels.

Implications for exposure assessment

Results from this and other studies27 32 33 34 35 36 are broadly consistent in the period 1999–2005 and can therefore be used to infer a power control factor of 0.25 to 0.5 for GSM network exposures in this time period. For any historical exposure index, however, extrapolation backwards in time would be necessary. GSM networks were introduced from the early- to mid-1990s in Europe and more recently in North America. Earlier, first-generation, analogue networks had no or very limited APC capability, so phones in early years were always operating at maximum power. Since the start of GSM systems, improved network coverage may have reduced power levels over time, but increased use of phones indoor, more frequent use of handovers, and higher traffic density may have increased them. Thus, there are considerable uncertainties about any backward extrapolation. Similarly, we cannot make extrapolations to new technologies with much lower output powers, such as recently introduced Universal Mobile Telecommunications Systems (UMTS) mobile phones. These are not relevant for historical exposure indices used in current studies (such as INTERPHONE), but will become important for future studies.

Although operator was an important determinant of output power in this study, we do not think our results can be used to estimate variation between operators in any exposure index, because of the limited temporal and spatial coverage of this study. We do recommend that exposure assessment in sparsely populated areas accounts for “urban-rural” differences. Usual duration of calls might also be used in constructing exposure indices, although self-reports of call duration will generally not be accurate enough for this purpose, as most of its effect on RF exposure was in calls of less than 30 s duration.

The INTERPHONE study has developed an exposure gradient combining information from this SMP study with information on amount and duration of mobile phone use, on manufacturers’ SAR measurements, and on the spatial SAR distribution inside the head,41 to estimate RF energy absorbed at the location of brain tumours.

Most epidemiological studies of mobile phone use have so far employed exposure assessments based only on questionnaires or interviews covering call frequency and duration of phone use, sometimes including the type of signal (analogue versus digital). Our study has measured power levels in GSM networks, which are substantially higher than the minimum levels theoretically achievable. Exposure indices could be improved by accounting for average power levels of different telecommunications systems. There appears to be little value in gathering information on circumstances of phone use apart from use in very sparsely populated regions.