Skin cancer and arsenic dermatoses (hyperkeratoses and dyspigmentation) have been observed in a three village area of Huhhot, Inner Mongolia, China, in 1991. Public health authorities have attributed the disease to arsenic in the drinking water based on their public health studies. They have conducted medical examinations of residents (n=3,228), obtained their well-use histories (n = 3179), and measured the arsenic content in local wells (n = 184). Other sources of arsenic have been sought but not found. This report presents the results of the analyses of their data. It is the first report examining the occurrence of skin cancer and arsenic dermatoses where exposure data is analyzed on an individual basis.

The analyses examine two arsenic exposure measures [peak arsenic concentration (PAC) exposure measure and a cumulative arsenic dosage (CAD) exposure measure] and four dermatological disorder outcome measures [hyperkeratoses, dyspigmentation, hyperkeratoses with dyspigmentation, and skin cancer]. Dose-response relationship analyses across exposure strata were conducted using frequency-weighted, simple-linear, hockey-stick, and most-likely estimate (MLE) models. Formal latency analyses were conducted using the MLE model.

The peak arsenic exposures ranged from undetectable (<10 ppb) to 2,000 ppb arsenic. 35% of the population had PAC measures below 50 ppb, 69% below 100 ppb, and 86% below 150 ppb. The cumulative arsenic dosages ranged from undetectable to 20372 ppb-years arsenic with 71% below 2,000 ppb-years exposure.

Eight skin cancer cases were identified as were 172 cases of hyperkeratoses, 121 cases of dyspigmentation, and 94 cases with both hyperkeratoses and dyspigmentation. All eight skin cancer cases occurred in individuals with both hyperkeratoses and dyspigmentation and with peak arsenic concentrations of 150 ppb or greater. Although cases of hyperkeratoses and of dyspigmentation occurred with peak arsenic concentrations under 50 ppb, they did not reach expected prevalences until higher peak arsenic concentrations.

These analyses reveal a statistically significant deficit of skin cancer among those with PAC exposures below 150 ppb, and particularly among those with PAC level between 50 ppb and 150 ppb. The dose-response curve for skin cancer is best described with respect to the peak arsenic concentration (PAC) by a frequency-weighted model with a threshold at or near 150 ppb arsenic or by a most likely estimate hockey-stick model with a threshold at 122 ppb arsenic. These results are consistent with the threshold-model analysis of the Taiwan data set that had showed a threshold at about 120 ppb. Issues of time consideration, latency, and misclassification have been considered, but do not at present appear to have markedly affected the analysis. A number of different approaches have been used to deal with confounding due to age, including the use of age-adjusted rates and of stratified analyses. This study should be replicated or expanded to further answer analytic questions.

In May 1990, the Sanitation and Anti-Epidemic Station (public health department) in Huhhot, Inner Mongolia (China) sought to determine why the public health clinic for the village of Zhi Ji Liang was requesting far more dermatological preparations than were other clinics. Examination of the dermatological cases of this village by the clinicians from the station led to the identification and diagnosis of chronic arsenicism cases (Luo Zhen Dong, MD; Chief Physician). A public health investigation of the extent of chronic arsenicism was carried out in this village and in two other villages in the area (Tie Men Geng and Hei He). The source of the arsenic was sought. The original findings and the public health investigation were published in the Chinese public health literature.

The investigations reached the conclusions that the wells providing the local water supply was the source of the arsenic exposure. Well-water samples were taken from the wells in the three villages. Arsenic levels above the public health standard of 50 ug/l were found in about two-thirds of the wells. Well-water was also tested for other water quality and inorganic measures. Other sources of arsenic were sought, including occupational, therapeutic, and dietary, but none were found. Environmental sampling included indoor and outdoor air, soil, and water. The results of the original clinical and environmental studies were presented at the Second International Conference on Arsenic and Health (San Diego, California; June 1995) and published in the English language proceedings and abstracts.

The arsenic contamination of the wells was found to come from natural sources. As the underground waters that were tapped by the wells were in a Q₄ earth stratum with a local rock having a high concentration of arsenic, the waters were found to have high levels of arsenic (mainly As⁺³). The Huhhot region sits in a triangular, segmented and sunken lake basin in Inner Mongolia, south of the Daqing (Great Green) Mountains and along the northern coast of the Yellow River. This region comprises approximately 4,800 km². The naturally arsenic-contaminated well-water represents an environmental and health problem in the areas surrounding Huhhot, the capital of Inner Mongolia, China. The chronic arsenicism investigation conducted in these areas by the Huhhot Sanitation and Anti-Epidemic Station began in 1990 and continues a decade later.

The original study upon which this analysis is based was designed to assess the prevalence and etiology of chronic arsenicism in three studied villages. The study population consisted of the total population of the three studied villages. Clinical examinations were conducted on each participant by physicians from the Huhhot Sanitation and Anti-Epidemic Station. Well-use histories were also obtained from each participant. Cases of chronic arsenicism were diagnosed on the basis both of clinical findings (hyperkeratosis and hyper/hypopigmentation) and environmental exposure history (more than six months exposure to a well with an arsenic concentration greater than 50 ug/l, the public health standard in China). Thus, the Chinese chronic arsenicism analyses of their outcome data were not exposure-independent.

The analytic study reported here has been conducted in order to separate out exposure and outcome information and then to examine the dose-response relationships between the individual dermatological findings (hyperkeratosis, dyspigmentation, and skin cancer) and the individual arsenic exposure histories of the participants. The previously published analysis was limited to an examination of those individuals exposed to arsenic at greater than 50 ug/l for at least six months as that exposure criteria was part of the case definition. The current study avoided the problem of confounding outcome with exposure by analyzing the clinical findings independently of the level of arsenic exposure and thus was able to extend the analysis well below the 50 ug/l level.

This analysis was based on a dataset originally assembled by Dr. Luo Zhen Dong and his colleagues at the Huhhot Sanitation and Anti-Epidemic Station. The analysis has been conducted collaboratively with the additional investigators of the Inner Mongolia Cooperative Arsenic Project (IMCAP). The Inner Mongolia Cooperative Arsenic Project was established as a collaborative research, analytic, and public health program between the Huhhot government (on behalf of the Huhhot Sanitation and Anti-Epidemic Station and the people of Huhhot) and the American collaborators (on behalf of themselves as arsenic health researchers and their institutions). IMCAP was established through a Memorandum of Cooperation jointly signed in Huhhot, Inner Mongolia in December 1994.

The arsenic health studies that had been conducted in Inner Mongolia were brought to the attention of the American collaborators in August 1994 by Professor C. J. Chen of the National Taiwan University School of Public Health Institute of Epidemiology. The American collaborators had previously voiced concern that risk assessments of cancer from the ingestion of arsenic had generally been based on the reported findings of Tseng et al. from the 1960s from Taiwan -- based on the data presented in their analytic tables and not on the underlying data, and using data from another country (Singapore) as their comparative reference population. These American scientists had cited a number of problems with using the Tseng study as the basis for the cancer risk assessment from arsenic ingestion. A principal problem is that the Taiwan study was an ecological study where average effects on people were compared with average arsenic concentrations. The scientists had argued that new data bases should be sought that would confirm the Taiwan observations and allow a dose-response analysis based on the primary data with comparative data from the same area. When they had become aware of the circumstances in Inner Mongolia, they sought to open up an opportunity to develop such analyses. Thus, IMCAP came to fruition.

The IMCAP analysts also sought to explore the pattern that might follow, if analysis were allowed to conform with a threshold model rather than default to a non-threshold model. Their previously published analysis of the data from the skin cancer-arsenic ingestion literature had suggested that the literature, particularly the Tseng et al. study, were consistent with the threshold model, although the non-threshold model was the model of choice of various regulatory organizations, including the US Environmental Protection Agency.⁵

This particular analysis has been undertaken for the Agency for Toxic Substance and Disease Registry under grant [# H75/ATH682885] to the University of Texas-Houston Medical School (Department of Dermatology) to examine epidemiologically the distribution of arsenic skin disease (including skin cancer) among persons exposed to arsenic water concentrations below 50 ug/l within the context of a population including subjects exposed to higher levels. This database has the distinct advantage that it has clinical outcome information and exposure data separately for each of approximately three thousand subjects.

The purpose of this analysis is to examine the relationship between chronic arsenic exposure through well water consumption in Inner Mongolia and selected dermatological disorders, including skin cancer. Two a priori hypotheses are examined in this report:

(a) Does the dose-response relationship between skin cancer and arsenic exposure follow a linear, non-threshold model (the standard cancer default model) or a linear or non-linear, threshold model (as suggested in Taiwan with a threshold of about 120 ppb)? In particular, ATSDR has asked whether the risk of skin cancer at exposures below 50 ppb arsenic is linearly predicted from the skin cancer risk at higher exposure levels.

(b) Are hyperkeratoses and dyspigmentation (as evidence of arsenic skin damage) necessary and sufficient pre-conditions for arsenic-induced skin cancer? Are the dose response relationships similar?

The three villages had a total population of 3,329 at the time of the surveillance (1992), of which 3,228 individuals participated in the study (participation rate = 97%). The high participation and surveillance rate reflects the agrarian nature of the villages and the lack of opportunity for out-migration. Both well-use history data and skin disease diagnostic data were recorded for almost all participants (3,179/3,228 = 98.5%). Well-use data was absent for 45 participants, and dermatological findings were absent for four participants. Eight individuals had well-use histories that included use of unmeasured wells. Their arsenic exposure estimates were based on their use of the wells with measured arsenic levels. The average well-use history was greater than 25 years. The demographic comparisons between the individuals with known exposure and outcome status and those with either unknown use or outcome status are tabulated in Table 2.1. Those with unknown exposure or outcome status were mostly children (92% less than 10 years old compared with 14% for those with known exposure and outcome). The numbers of male and female subjects in the study population were similar. Almost all study subjects identified themselves as being of Han (H) (Chinese) origin rather than of Mongolian (M) origin.

Table 2.1. Distribution of individuals with and without known exposure and outcome status

E/O                  # of Subj.       Age  (SD)                   Gender                          Race
Status                      (N)         (years)              M (%)           F (%)         H (%)            M (%)

Known                3179              33 (20)       1628 (51)     1551(49)   3172 (99.8)      7 (0.2)
Unknown                49               7 (10)            22 (45)          27 (55)     49 (100.0)     0 (0.0)

Total                    3228             32 (20)       1650 (51)      1578 (49)   3221 (99.8)     7 (0.2)

Subsequent analysis was limited to those with both known exposure and outcome status. Analysis of the demographic distributions of the three selected villages (Table 2.2) showed that people who lived in Hei He tended to be older than those in Tie Men Geng and Zhi Ji Liang, though the gender proportions were similar among the three villages. The seven individuals of Mongolian ethnic origin lived in Hei He. All subjects with a known occupation were either students or farmers, whether male or female.

Table 2.2. Comparison of demographic distribution among the three selected villages

                            # of Subj.      Age (SD)                      Gender                          Race
Village                      (N)            (years)            M (%)         F (%)          H (%)            M (%)

Hei He                     1755           36(19)          901 (51)      854 (49)     1748 (99.6)    7 (0.4)
Tie Men Geng            257           27(19)          128 (50)      129 (50)       257 (100)     0 (0.0)
Zhi Ji Liang              1167            30(19)          599 (51)      568 (49)    1167 (100)      0 (0.0)

The frequency distribution of the study population (n=3,179) by age group and village are shown (Table 2.3). Age was missing for four individuals (one from Tie Men Geng and three from Hei He). The median age was 29 years.

Table 2.3. Distribution of the study population by age group and village
 

Age (yrs)	Hei He		Tie Men Geng		Zhi Ji Liang		Total
	n	%	n	%	n	%	n	%
<10	117	6.7	64	25.0	215	18.4	396	12.5
10-19	326	18.6	38	14.8	191	16.4	555	17.5
20-29	346	19.8	49	19.1	262	22.5	657	20.7
30-39	308	17.6	53	20.7	194	16.6	555	17.5
40-49	259	14.8	12	4.7	92	7.9	363	11.4
50-59	164	9.4	21	8.2	100	8.6	285	9.0
60-69	130	7.4	11	4.3	78	6.7	219	6.9
70+	102	5.8	8	3.1	35	3.0	145	4.6
Total	1752	100.0	256	100.0	1167	100.0	3175	100.0

Well water samples were collected by the Huhhot Sanitation and Anti-Epidemic Station and analyzed in their laboratory using a silver diethyl-dithiocarbamate colorimetric methodology with a 10 ug/l detection limit.^8,10 Laboratory quality control was supervised by the Chinese Academy of Preventive Medicine Laboratory of Environmental Engineering. Split samples have been analyzed with good agreement, and replicate measurements at the National Taiwan University laboratory have confirmed the reliability of the measurements. Arsenic measurements were available for 184 of the 187 wells mentioned in the well-use histories. The three other wells had been closed in 1957-1959 and were not available for testing. Use of these wells was reported by only eight participants and had occurred beginning as early as 1920. No surrogate measurements were used for wells that could not be sampled. The number of wells varied among the three villages - 79 for Hei He, 45 for Tie Men Geng, and 60 for Zhi Ji Liang. The number of water samples taken from each of these wells also varied, with one sample taken from 165 wells, two samples taken from 18 wells, and three samples taken from one well (detailed by village in Table 3.1). When more than one sample was taken from a single well, the geometric mean of the repeated measures of the As concentration was used. The As concentrations for the 184 wells varied widely from non-detectable (<10 ppb) to 2,000 ppb. The frequency distribution of the wells by villages, by the number of sample taken, and by As concentration groups are presented in Table 3.1. The descriptive statistics of arsenic concentration by village are also displayed in Table 3.1.

Table 3.1. Frequency distribution of wells and descriptive statistics of As concentration



                         Frq. by sample #                        Frq. by As concentration grps                 As concentration statistics (ppb)
Village           One   Two     Three             <10    10-     50-     150-       500+              Mn     Std.    Min P25    Med     P75       Max

Hei He       76     2         1               1   29    46     3         0             65     35     5     37    57    81     176
Tie MG      34    11        0               8   16    16    4         1             75   106     5     20    48    80     615
Zhi JL         55     5         0             35   13     1     4         7            156  386     5      5      5     27    2000

Total         165   18         1             44   58   63    11        8             97    230    5     14    43    80   2000

Whether well water arsenic levels varied significantly over time was unclear. Most of the wells had only a single arsenic measurement. Those measurements were assumed to be representative of those wells over a long period of time. The data on the 19 wells with more than one measurement allowed us to examine their variation over time. The intervals between the first and second measurements for the 19 wells ranged from 1 to 6 years with a mean of 3 ¼ years. The distributions of paired measurements were examined (Figure 1). The sample distribution within each set had the same pattern. The means of the two distributions did not significantly differ in a paired t-test analysis (p > 0.05; Pearson correlation = 0.89). As the data on the wells with replicate analyses showed no significant variability on replicate sampling over time, it is a reasonable assumption with respect to the geological circumstances in Inner Mongolia, that the same lack of variation applied to the remaining 90 % of the wells, for which we had only one sample measurement over the time period for which we have no data.

The original Chinese studies used a minimum exposure duration of 6 months in their reports. This report is based on exposures reported as years of exposure, so the minimum exposure time is 12 months.

3.2   Measurement of exposure

Arsenic exposure of the subjects were analyzed using two different measures, the peak arsenic concentration (PAC, in ppb) of the well waters ever consumed and the cumulative arsenic dosage (CAD, in ppb-year) determined from the individual’s well-use history. The well-use histories of the participants included as many as five different wells for a single individual. The individual’s complete well-use history while resident in these villages was utilized in calculating the cumulative dosage. The cumulative dosage was calculated using the following formula: CAD = sum of (As concentration Xexposure years) for each well use. For purposes of calculation, the samples with non-detectable As levels were set at 5 ppb, half of the detection limit. The descriptive statistics of the peak As concentration and lifetime cumulative dosage were calculated and are displayed in Table 3.2. The numbers in the two groups differ, because the well use history of one individual identified the wells used but not the time periods. Thus, a peak arsenic concentration could be calculated but not a cumulative arsenic dosage.

Table 3.2. Descriptive statistics of peak As concentration and cumulative As dosage

Exposure           N        A-Mean      A-Std.   G-Mean   G-Std    Min         P25        Med        P75         Max

PeakCon        3179      85             134           50              3.0     < 10        30        58           110            2000
CumDose       3178    1571         1982         740              4.2     < 10      300       948          2184        20372

No data exists on the daily water consumption rate of the participants from these villages. However, as all three villages are similar agrarian communities in close proximity to each other, it is assumed that the water consumption rates in the villages are similar. This should not affect the risk analyses when exposures are reported as either ppb arsenic or ppb-years arsenic. Such an estimate would, however, if risk estimates are to be developed using exposures reported as either milligrams of arsenic per day or cumulative arsenic exposure in milligrams or grams.

3.3 Categorization of exposure

Table 3.3. Descriptive statistics for the eight PAC groups (ppb)

PAC (ppb)             N     A-mean     A-SD      G-MN     G-SD     Min     P25     Med     P75      Max

<10                       287       5.0          0.0            5.0           1.0        5        5          5          5            5
10-                       405       14.7         4.7          14.1          1.3        10      13        13        17          28
30-                        412      32.8         3.3           32.6          1.1       30      30         34       34          49
50-                        516      54.9         3.1           54.8          1.1       50      52         56       58           59
60-                        565      70.3         9.4           69.7          1.1       60      62         62        81          92
100-                      536     121.7       18.6        120.3          1.2      100    105      115      147        147
150-                      416     174.6       32.8        172.5          1.2      156    158      158      182         345
500+                       42    1048.3     423.2       974.9          1.5      580    615      961    1174        2000

As the study population by the cumulative arsenic dosage (CAD) was approximately log normally distributed, the study population was categorized into eight exposure dose groups of equal intervals on the logarithmic scale (0.5 log units). The descriptive statistics for the eight CAD exposure groups are displayed in Table 3.4. The arithmetic means, the geometric means, and the medians for each CAD group were compared. Due to their similarity (see Table 3.4 and Figure 3), the arithmetic mean was used as the CAD type of exposure measure in the group analyses.

Table 3.4. Descriptive statistics for the eight CAD groups (ppb-year)

CAD                       N     A-mean      A-SD      G-MN     G-SD      Min      P25     Med     P75      Max

<10                          7         5.0            0.0          5.0             1.0         5         5          5          5           5
10-                          97         20            7.2            9              1.5        10        15        20        26        31
32-                        180         62             20           59             1.4        34        45        60        80        98
100-                      554       203             65         192             1.4       100      150     196      260       315
316-                      800        617          198         585             1.4       318      442      590      780      999
1000-                   1123     1893          603        1797            1.4     1000    1352    1856    2357    3154
3162-                     380     4656        1507        4462            1.3     3163    3590    4130    5254    9761
10000+                    37    12971        2425      12773           1.2   10059   10955 12740   14912  20372

The cumulative percentage of the study population was plotted linearly against the mean of the arsenic dosage intervals both for peak arsenic concentration (Figure 4.1) and for cumulative arsenic dosage (Figures 4.2). The study population had predominantly lower exposure levels for both exposures (Figures 4.1 and 4.2). The PAC ranged from 5 (i.e., non-detect) to 2000 ppb with 69% of the study population having a PAC of less than 100 ppb. The CAD ranged from 5 to 20,372 ppb-yrs with 71% of the study population having a CAD of less than 2,000 ppb-years.

Alternative sources of arsenic have been sought. The western Huhhot basin is an agricultural area, raising wheat, millet, corn, green beets, potatoes and sunflower. Arsenical pesticides have not been used. No factories, mines or other industries discharge arsenic into the local air, water, or soil. Examination of the surface soils, air, fish and crops have not found arsenic levels above those of the general Chinese culture. The smoking habits in Huhhot resemble those of the general Chinese culture.¹ The use of coal for household heating and cooking is another potential source of arsenic exposure. Its use, however, is similar across the three villages.

Three dermatological disorders were recorded in the original data set, hyperkeratoses, dyspigmentation (hyperpigmentation/hypopigmentation of the trunk) and skin cancer. The skin disorders were diagnosed by the Chinese physicians conducting the survey, using their established clinical criteria^1,2 for hyperkeratoses, dyspigmentation, and skin cancer. The term hyperkeratoses in this report referred to obvious thickening of skin on the palms and soles in palpable or wart-like bumps ranging in size from about approximately 0.2 to 1.5 cm over large areas, whether separated or coalesced. Dyspigmentation referred to coarse skin with moderately-sized spots of pigmentation, distributed in a web-like form. The diagnosis of dyspigmentation was made on the basis of truncal findings, rather than findings on the extremities, as they were less likely to be confounded by solar (actinic) exposure. The clinical skin cancer diagnoses were independently substantiated clinically and histologically (basal cell and squamous cell carcinomas) by Professor Stephen B. Tucker of the University of Texas Department of Dermatology. The non-malignant cutaneous findings reported by the Huhhot Sanitation and Anti-Epidemic Station physicians were also verified by Dr. Tucker and his co-workers.

The analyses here have been conducted for two non-malignant outcomes commonly attributed to arsenic exposure^1,2,3.4 (i.e., (a) hyperkeratoses; (b) dyspigmentation) and their co-appearance (hyperkeratoses and dyspigmentation) and for the malignant outcome of skin cancer. Hyperkeratoses were the most prevalent skin disease in the study population (5.4%), and dyspigmentation was second (3.8%). Combined hyperkeratoses and dyspigmentation had a prevalence of 3.0%. Skin cancer was the dermatological finding with the lowest prevalence (0.3%). The prevalence of hyperkeratoses without dyspigmentation can be calculated in each strata from the difference between the prevalence of hyperkeratoses and the prevalence of hyperkeratoses with dyspigmentation. The analogous statement holds for the prevalence of dyspigmentation without hyperkeratoses. Additionally, all eight study subjects with skin cancer had both hyperkeratoses and dyspigmentation.

Table 4.1 shows the number and prevalence of each type of skin disorder and the prevalence of skin cancer among those with the various disorders. The first two data columns show the number and prevalence of each skin condition within the total study population. The second two data columns show the number and prevalence of skin cancer within each dermatopathology group.

Table 4.1 Prevalences of Skin Disorders in the studied subjects and prevalence of skin cancer in each skin disorder group.

                                                                        All Subjects                                         Skin Cancer Cases
Skin Disorder                                                      N %*                                                        N %**

Total population                                              3179   100.0%                                              8    0.25%
No arsenic dermatosis                                     2980   93.7%                                                0    0.00%
Any arsenic dermatosis                                      199   6.3%                                                  8    4.02%
Hyperkeratoses (K)                                           172   5.4%                                                 8    4.65%
Dyspigmentation (P)                                           121   3.8%                                                 8    6.61%

Both (K) and (P)                                                 94    3.0%                                                 8    8.51%

Skin cancer                                                         8      0.3%                                                 8   100.0%

It is noteworthy that skin cancer was observed in one-twelfth (8.5%) of the subjects with both hyperkeratoses and dyspigmentation. All cases of skin cancer occurred in persons with both hyperkeratoses and dyspigmentation; However, most persons (92%) with both hyperkeratoses and dyspigmentation did not develop skin cancer.

5.0 Relationship between Exposures measured as PAC and Outcomes:

The relationship between peak arsenic concentration and each of the four outcomes was examined using a frequency-weighted model. The results are presented in Table 5.1 and Figures 5.1a – 5.1d.

As shown in Table 5.1, a general monotonically increasing dose-response pattern generally was seen for all four outcomes. A linear trend in proportion test that assigned each strata to its mean PAC exposure was highly significant for each of the four examined outcomes (all p < 0.01). Keratoses and dyspigmentation or both combined were observed in all exposure groups. In contrast, skin cancer cases were observed only in the two highest arsenic concentration groups, giving the impression of a threshold effect of arsenic concentration on skin cancer at a level of about 150 ppb.

An expected number of cases was calculated for each outcome and for each exposure stratum assuming (a) the total number of cases expected in the population was equal to the total number observed, (b) the background risks were too low to expect any background cases in a population of this size, and (c) the proportion of cases expected in each exposure stratum was the same as the proportion of the total exposure observed in each exposure stratum. The formula used for the calculation was:
 
 

N_expected = (N_i ´ X_i) / [å_i=1⁸ (N_i ´ X_i)]  ´ N_t
 

N_i: the number of subjects in the ith strata
X_i: the mean of PAC intervals in ith strata
N_t:the total number of cases observed

Table 5.1 shows the relationship between peak arsenic concentration and the four skin conditions.

Table 5.1. Relationship between peak arsenic concentration and the four skin disorders
 

PAC Group	Mean PAC (ppb)	Subj N	Exposure (ppb-person)	Keratoses (a)				Dyspigmentation (b)				Kera+Dysp(c)				Skin Cancer (d)
				Observed		Expected		Observed		Expected		Observed		Expected		Observed		Expected
				N	Prev	n	Prev	n	Prev	n	Prev	N	Prev	n	Prev	n	Prev	n	Prev
<10	5	287	1,435	1	0.3	0.9	0.3	3	1.0	0.6	0.2	1	0.3	0.5	0.2	0	0.0	0.0	0.0
10-	15	405	6,075	1	0.2	3.9	1.0	0	0.0	2.7	0.7	0	0.0	2.1	0.5	0	0.0	0.2	0.0
30-	33	412	13,596	4	1.0	8.6	2.1	4	1.0	6.1	1.5	3	0.7	4.7	1.1	0	0.0	0.4	0.1
50-	55	516	28,380	8	1.6	18	3.5	7	1.4	13	2.5	6	1.2	9.8	1.9	0	0.0	0.8	0.2
60-	70	565	39,550	23	4.1	25	4.4	11	1.9	18	3.1	11	1.9	14	2.4	0	0.0	1.2	0.2
100-	122	536	65,392	61	11	42	7.7	42	7.8	29	5.4	31	5.8	23	4.2	0	0.0	1.9	0.4
150-	175	416	72,800	45	11	46	11	34	8.2	33	7.8	24	5.8	25	6.1	5	1.2	2.1	0.5
500+	1048	42	44,016	29	69	28	67	20	48	20	47	18	43	15	36	3	7.1	1.3	3.1
Total	85	3,179	271,244	172	5.4	172	5.4	121	3.8	121	3.8	94	3.0	94	3.0	8	0.3	8.0	0.3

Note:
PAC Group: Exposure groups based on peak arsenic concentration
Mean PAC: Means of the peak arsenic concentration for each exposure group
n: number of cases
Prev: Prevalence as cases per 100 subjects

Visual examination of Table 5.1 and Figures 5.1a – 5.1d gives the impression that the non-malignant skin disorders (keratoses, dyspigmentation, or both) are predicted better by the frequency-weighted model than is skin cancer. Skin cancer tends to be over-predicted for exposure below 150 ppb and to be under-predicted for exposure above 150 ppb. A comparison of the observed and the expected number of cases found no significant difference either for the non-malignant skin disorders ( C²_df=7=11, p=0.13 for keratoses; C²_df=7=10, p=0.17 for dyspigmentation; C²_df=7=5.3, p=0.62 for both combined) or for skin cancer ( C²_df=4 = 6.3, p = 0.18). The number of observed lesions lies below the expectation for the low PAC groups. This suggests the possibility that better fits may be obtained for models that permit a threshold or other sub-linear dose-response relationship.

The prevalences (P) of the four skin disorders were fitted to a simple function of the means of the PAC intervals (P = a+ b*Dose) using an unconstrained least squares linear model with each bin weighted equally. a was not constrained to be positive. The fitted parameters of the model for each of the four outcomes are presented in Table 5.2. The observed and predicated prevalences by the means of the PAC intervals for each of the four skin disorders are displayed in Figure 5.2a – 5.2d. The p-value presented in the 6^th column is the probability that there is no true dose dependence (b = 0) and that the events are random samples. The p value presented below the table is the probability that the true threshold is zero and the observed value of 43 ppb occurs by chance.

Table 5.2. The parameters of simple linear modeling for PAC exposures
 

	Skin Disorders	Y-Intercept (a)	Slope ( b)	R 2	F(1,6)	p-value	Unit Risk (b )	X-Intercept (-a / b)
a	Keratoses	-0.0034	0.00066	0.995	1300	3.0E-08	6.6E-04/ppb	4.9 ppb
b	Dyspigmentation	-0.00073	0.00046	0.995	1165	4.2E-08	4.6E-04/ppb	1.4 ppb
c	Kera+Dysp	-0.0055	0.00041	0.998	2611	3.8E-09	4.1E-04/ppb	14 ppb
d	Skin Cancer	-0.0030	0.00007	0.988	473	6.2E-07	0.71E-05/ppb	43 ppb*

* p<0.05

As shown in Table 5.2, the four simple linear regressions are all statistically significant. The measure of arsenic contamination (i.e., the mean of the PAC intervals) explains about 99% of the overall variation of prevalence for the four skin disorders. The unit risk (change of the prevalence with each unit change of the mean of the PAC interval), or the slope calculated by this model, is similar for the non-malignant skin disorders and is about an order of magnitude higher than the slope for skin cancer.

In order to determine whether the relationships between the exposure to arsenic and the four outcomes are consistent with a non-threshold linear model or with a threshold linear model, the x-intercept (-a / b) of each fitted linear model was determined. The x-intercept was used as the best estimate of the potential PAC threshold value for each outcome examined (see Table 5.2 and Figure 5.2a – 5.2d). The 95% confidence intervals for the x-intercepts were calculated from the 95% confidence intervals of the y-intercept (a ) derived using a commercial software program, STATA. As shown in Table 5.2, skin cancer has the largest x-intercept (43 ppb with 95% CI 0.4 – 96) as compared to keratoses (4.9 ppb with 95% CI -19 – 33) or dyspigmentation (1.4 ppb with 95% CI -24 – 31), or both non-cancer skin lesions combined (14 ppb with 95% CI -4.8 – 33). In those circumstances in which the 95% confidence intervals of the x-interval were found to exclude the value of 0 (zero), the data were inconsistent with the non-threshold linear model and consistent with the threshold linear model.

Since the simple function allows P (prevalence) to be negative and to be greater than unity (although no data can do so), a fit to a more realistic formula was made. A program kindly provided to us by Edmund A. Crouch, PhD enabled us to use a hockey-stick model modified to ensure that P reached, but did not exceed, unity at high doses. The hockey-stick model alsoallowed for the possibility of a non-zero intercept (threshold) in the dose-response relationship and terms involving higher powers of the dose than the linear term. The prevalence of the four skin disorders was fitted to this hockey-stick model using a maximum likelihood method.

The precise description of this model is P = 1-exp(-a) for dose (d) less than a threshold dose (d_t), and P = 1-exp( - ( a + b*(d-d_t)) for d > d_t where a (alpha) and b (beta) are constrained to be positive, although this constraint does not appreciably affect the derived parameters. 1-exp( -a) ~ a is the "background" of the lesion at zero dose. The parameters of these model fits for each of the four outcomes are presented in Table 5.3. The observed and predicted prevalence by the means of the PAC intervals for each of the four skin disorders are displayed in Figure 5.3a –5.3d.

Table 5.3. The parameters of the hockey-stick model fits for PAC exposure measures
 

	Skin Disorders	a	b	C2(df=5)	GOF test p	Threshold (dt)
a	Keratoses	0.0055 0.0029	0.00110 0.00097	11 11	0.06 0.06	42 ppb* 30 ppb*
b	Dyspigmentation	0.0101 0.0080	0.00071 0.00072	9.0 8.5	0.11 0.13	50 ppb* 47 ppb*
c	Kera+Dysp	0.0057	0.00052	5.8	0.33	42 ppb*
d	Skin Cancer	0.00000	0.00015	2.2	0.91	122 ppb*

The range of uncertainty for the threshold was found by plotting C² values against the assumed threshold achieved when the model parameters were readjusted to get the best fit. The X ² value is increased above the minimum value (shown in Table 5.3) by +2 for the two threshold dose values that differ from the best value by two standard deviations (approximately the 95% confidence intervals) and by +1 for one standard deviation (not shown).

The threshold values (d_t) for the hockey-stick models in Table 5.3 are greater than the X-intercepts for the simple linear fit models in Table 5.2. That is because the program does not have to try unsuccessfully to fit the zero lesions at doses below the threshold. All the fits were acceptable (p > 0.05), but those for keratoses and dyspigmentation were not very good. Addition of extra parameters can usually improve a fit, but when extra terms with a power of the dose greater than one were added to the model, the coefficients of these terms were zero and the goodness-of-fit was only slightly improved. As is usual in fits to cancer models, the coefficients were constrained to be positive. Not all analysts regard tests of higher powers within a hockey-stick model to be a reliable test of the existence of a threshold. Nonetheless, neither a quadratic nor a cubic or higher term would improve these fits. This is in contrast to the simple fits to the Taiwan skin cancer ecological data where a cubic term obviated the need for a threshold.

Although none of the skin cancers were observed at exposure levels below the calculated skin cancer threshold level, some cases of keratoses and/or dyspigmentation are reported at levels below their calculated threshold values. Whether these cases represent diagnostic errors, exposure mis-classifications, background cases, or evidence of a relative threshold in dose-dependency rather than an absolute one is uncertain.

The relationships between cumulative arsenic dosage and each of the four skin disorders were also examined. As with the peak arsenic concentration, a general dose-prevalence pattern (higher prevalence for higher cumulative As dosage group) was also seen for all four skin disorders. The linear trend in proportion test was highly significant for each of the four outcomes examined (all p < 0.01). Skin cancer cases occurred only in the three highest cumulative arsenic dosage groups (>1000 ppb-years), also suggesting a threshold for cumulative exposure to arsenic on skin cancer, but consistent with both a threshold and a non-threshold linear model.

The expected number of cases for the four outcomes was calculated using the same formula for CAD exposures as for PAC exposures (N_expected = ((N_i x X_i)/ å_i=1⁸ (N_i x X_i)) xN_t). As shown in Table 6.1 and Figure 6.1a – 6.1d, the frequency-weighted approach predicts both non-malignant skin disorders and skin cancer, with some over- and under- predicting.

Table 6.1. Relationship between cumulative arsenic dosage and the four skin disorders
 

CAD Group	Mean CAD (ppb-yr)	Subj N	Exposure (ppb-person-yr	Keratoses (a)				Dyspigmentation (b)				Kera+Dysp (c)				Skin Cancer (d)
				Observed		Expected		Observed		Expected		Observed		Expected		Observed		Expected
				n	Prev	n	Prev	n	Prev	n	Prev	n	Prev	n	Prev	n	Prev	N	Prev
<10	5	7	35	0	0.0	0.0	0.0	0	0.0	0.0	0.0	0	0.0	0.0	0.0	0	0.0	0.0	0.0
10-	20	97	1,940	0	0.0	0.1	0.1	0	0.0	0.0	0.0	0	0.0	0.0	0.0	0	0.0	0.0	0.0
32-	62	180	11,160	0	0.0	0.4	0.2	1	0.6	0.3	0.2	0	0.0	0.2	0.1	0	0.0	0.0	0.0
100-	203	554	112,462	2	0.4	3.9	0.7	2	0.4	2.7	0.5	1	0.2	2.1	0.4	0	0.0	0.2	0.0
316-	617	800	493,600	12	1.5	17	2.1	7	0.9	12	1.5	5	0.6	9.3	1.2	0	0.0	0.8	0.1
1000-	1893	1123	2,125,839	77	6.9	73	6.5	56	5.0	52	4.6	44	3.9	40	3.6	4	0.4	3.4	0.3
3162-	4656	380	1,769,280	64	17	61	16	43	11	43	11	35	9.2	33	8.8	3	0.8	2.8	0.7
10000+	12971	37	479,927	17	46	16.5	44.7	12	32	12	31	9	24	9.0	24	1	2.7	0.8	2.1
Total	1571	3,178	4,994,243	172	5.4	172	5.4	121	3.8	121	3.8	94	3.0	94	3.0	8	0.3	8	0.3

Note:
CAD Group: Exposure groups based on both peak arsenic concentration and exposure duration, i..e., cumulative arsenic dosage
Mean CAD: Means of the cumulative arsenic dosage for each exposure group
Prev: Prevalence

A comparison of the observed and the expected number of cases showed no significant difference, either for the non-malignant skin disorders ( C²_df=7=1.7, p=0.78 for keratoses; C²_df=7=1.7, p=0.80 for dyspigmentation; C²_df=7=1.7, p=0.79 for both combined) or for skin cancer ( C²_df=4=1.1, p =0.77).

As with the PAC exposure measure, the prevalence of the four skin disorders was also linearly fitted against the mean of the CAD intervals. The parameters of the least squares fit for each of the four outcomes are presented in Table 6.2. The observed and predicted prevalences by the means of the PAC intervals for each of the four skin disorders are displayed in Figure 6.2a –6.2d. The p-values represent the likelihood that the slope is no different from zero.

Table 6.2. The parameters of the simple linear model for CAD exposures
 

	Skin Disorders	Y-Intercept (a)	Slope (b )	R 2	F(1,6)	p-value	Unit Risk (b )	X-Intercept (-a / b)
a	Keratoses	-0.0017	0.000036	0.999	17720	1.2E-11	3.6E-05/ppb-yr	43 ppb-yr
b	Dyspigmentation	-0.00065	0.000025	0.999	6767	2.2E-10	2.5E-05/ppb-yr	35 ppb-yr
c	Kera+Dysp	-0.00053	0.000019	0.999	4627	6.8E-10	1.9E-05/ppb-yr	25 ppb-yr
d	Skin Cancer	-0.00051	0.000002	0.995	1168	4.2E-08	2.1E-06/ppb-yr	313 ppb-yr

All four simple linear regression analyses are highly significant (Table 6.2), that is that the slope is clearly different from zero. The cumulative type of exposure to arsenic in drinking water can explain 99% or more of the overall variation of prevalence within the cumulative exposure groups for the four skin disorders. Again, the unit risk is similar for keratoses and dyspigmentation or both combined, while skin cancer has the lowest unit risk (about an order of magnitude lower). As with PAC exposures, skin cancer has the largest x-intercept (313 ppb-yrs with 95% CI -100 – 641), about an order of magnitude greater than keratoses (43 ppb-yrs with -42 – 141) and dyspigmentation (35 ppb-yrs with 95% CI -115 – 175), or both combined (25 ppb-yrs with 95% CI -143 – 211). As none of these x-intercepts are significantly different from zero, these analyses do not show evidence of a threshold.

Since cumulative measures of exposure include time, in addition to well water concentration, PAC and CAD exposures are not necessarily comparable. So, these analyses are not in conflict. The toxicological concept of a threshold implicit in an acceptable daily intake estimate does not include duration of exposure.

6.3. Hockey-stick model

The prevalences of the four skin disorders were also fitted as functions of both the means of CAD intervals and potential thresholds, using the same hockey-stick model and procedures used for the PAC exposures. The parameters of the hockey-stick model for each of the four outcomes are presented in Table 6.3. The observed and predicted prevalences by the means of the CAD intervals for each of the four skin disorders are displayed in Figures 6.3a – 6.3d.

Table 6.3. The parameters of the hockey-stick model for CAD exposures
 

	Skin Disorders	a	b	C2(df=5)	GOF p test	Threshold (dt)
a	Keratoses	0.0036 0.0000	0.000043 0.000040	1.3 1.4	0.93 0.97	353 ppb-yrs* 135 ppb-yrs*
b	Dyspigmentation	0.0041	0.000030	1.2	0.95	440 ppb-yrs*
c	Kera+Dysp	0.0018	0.000024	1.0	0.96	406 ppb-yrs*
d	Skin Cancer	0.00000	0.000002	0.2	0.9998	617 ppb-yrs+

*+ Significantly different from zero at * p < 0.05 + p < 0.10

Table 6.3 presents the regression coefficients and the test of goodness-of-fit. As indicated by the p values for the goodness-of-fit tests, the fits are very good (p for goodness-of-fit test > 0.05 and close to unity) for all the four disorders with statistically significant threshold values for the three non-malignant disorders [keratoses (353 ppb-yrs, 95% CI 213 – 459 and 135 ppb-yrs, 95% CI 50-172), dyspigmentation (440 ppb-yrs, 95% CI 263 - 560), or both combined (406 ppb-yrs, 95% CI 206 - 532)]. The skin cancer data analysis revealed three minima including one at 617 ppb-yrs that was significant on a one-tailed test (95% CI 179-1019) but not a two-tailed test (95% CI –33 - +1168). These analyses did not include models anchored at any well water concentrations above zero. Therefore, the analyses cannot exclude non-zero thresholds.

Age is a confounder of risk from a chronic exposure. Adjustment of the crude rate by age stratification to a standard population distribution can adjust for the degree of confounding contributed by the differences in age structure of comparative populations. Both crude prevalence rates and age-adjusted prevalence rates for the various dermatological conditions when examined by strata of the PAC exposure are shown in Table 7.1. The age-adjusted prevalence rates for the various dermatological conditions stratified by the PAC exposure differ little from the crude rates, suggesting no major confounding by age for the PAC exposure. Graphic comparison of the crude and age-distributed dermatological prevalence rate by PAC exposure as seen in Figure 7.1.

Table 7.1 Crude and (age-adjusted) Dermatological Prevalence Rates by PAC Exposure
 

PAC	Keratoses	Dyspigmentation	Kerato/Dyspig	Skin Cancer
< 10	0.4 (0.4)	1.1 (1.1)	0.4 (0.4)	0.0 (0.0)
10-	0.6 (0.6)	0.5 (0.6)	0.4 (0.4)	0.0 (0.0)
50-	5.7 (5.4)	3.7 (3.5)	3.0 (2.7)	0.0 (0.0)
150-	11 (9.4)	8.2 (7.0)	5.8 (4.8)	1.2 (1.0)
500+	69 (71.9)	48 (53.7)	43 (48)	7.1 (5.9)

 

Similarly, Table 7.2 presents the crude and age-adjusted prevalence rates for the various dermatological conditions when examined by strata of the CAD exposure. The prevalence rates within each exposure bracket have been adjusted in order to make the observations comparable. It is predictable that the duration of exposure and the age should co-vary, thus it is most likely that the cumulative exposure and the age will co-vary. It is observed that in the CAD analysis the age-adjusted prevalence rates for the non-malignant skin conditions tended to be greater than the crude rates, while the age-adjusted prevalence rates for the malignant skin condition tended to be lower than the crude rates. Graphic comparison of the crude and age-distributed dermatological prevalence rate by CAD exposure as seen in Figure 7.2.

Table 7.2 Crude and (age-adjusted) Dermatological Prevalence Rates by CAD Exposure
 

CAD	Keratoses	Dyspigmentation	Kerato/Dyspig	Skin Cancer
<10	0.0 (0.0)	0.0 (0.0)	0.0 (0.0)	0.0 (0.0)
10-	0.0 (0.0)	0.0 (0.0)	0.0 (0.0)	0.0 (0.0)
32-	0.0 (0.0)	0.6 (0.5)	0.0 (0.0)	0.0 (0.0)
100-	0.4 (0.4)	0.4 (0.4)	0.2 (0.2)	0.0 (0.0)
316-	1.5 (2.0)	0.9 (1.2)	0.6 (1.0)	0.0 (0.0)
1000-	6.9 (5.8)	5.0 (4.3)	3.9 (3.0)	0.4 (0.3)
3162-	17 (23)	11 (16)	9.2 (14)	0.8 (0.2)
10000+	46 (54)	32 (53)	24 (39)	2.7 (2.0)

In this report, the relationship between chronic arsenic exposure and four skin conditions (hyperkeratoses, dyspigmentation, hyperkeratoses with dyspigmentation, and skin cancer) was examined, using 3,179 residents of three villages in Huhhot, Inner Mongolia, China. The arsenic exposures arose primarily from the consumption of local arsenic-contaminated well water. The exposures were estimated, both using their peak arsenic concentration (ranging from non-detect to 2,000 ppb As) and using their cumulative arsenic dosage (ranging from non-detect to 20,372 ppb-yrs). A monotonically increasing exposure-response pattern was generally found for all four skin conditions and for both types of exposures.

A frequency-weighted model was used to determine the expected number of cases of skin cancer in each exposure strata, assuming that the prevalence of skin cancer was directly proportional to the exposure dose. The observed cumulative skin cancer case count by peak arsenic concentration and that expected on a linear model are presented in Table 8.1 and Figure 8.1.

Table 8.1. Observed and expected cumulative skin cancer case count by peak arsenic concentration (frequency weighted model)
 

PAC groups  (ppb)	Mean  PAC (ppb)	Cumulative  Subject. N	Cumulative Observed N	Cumulative Expected N
<10	5	287	0	0.04
10-	15	692	0	0.22
30-	33	1104	0	0.62
50-	55	1620	0	1.46
60-	70	2185	0	2.63
100-	122	2721	0	4.55
150-	175	3137	5	6.70
500+	1048	3179	8	8.00

Table 8.1 shows that none (0) of the skin cancer cases were observed at PAC exposures below 150 ppb, whereas 4.55 of the cases were expected in that range. The frequency-weighted linear model has significantly over-predicted the distribution of observed skin cancer cases for peak arsenic concentrations below 150 ppb (p= 0.02). This analysis demonstrates that the linear no-threshold model does not fit the data, as it significantly over-predicts the case count below 150 ppb. The criteria that the absence of cases (i.e., n = 0) is a statistically significant observation is that the expected number of cases for that group must be 3.7 or greater (Poisson two-tailed, p < 0.05), which it is. Thus, this study was sufficiently large t