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 to make that statement for exposures under 150 ppb As.

Table 8.3 below shows that the mean ages of the subjects in the PAC exposure groups below 50 ppb range between 26 and 30 years and that the mean ages of the subjects in the PAC exposure groups of 50 ppb and above range between 32 and 37 years. Thus, those at the lower PAC exposures (i.e., < 50 ppb) tended to be younger than those with the greater PAC exposures (i.e., 50 ppb and above). This age confounding is greatly limited by examining the rates in those in the 50 ppb and above group. It is now noted that none of the skin cancer cases were observed at PAC exposures of 50-149 ppb, where as 3.93 of the cases (4.55-0.62 = 3.93) were expected in that range. These data demonstrate a statistically significant deficit of skin cancers observed in the 50-149 ppb arsenic PAC exposure range.

To make a similar statement for the below 50 ppb As range, the study would need to be much larger. The study would have to be six times greater (6 x 0.62 > 3.7) if that population had the same age-distribution as those with 50 (+) ppb arsenic. This would involve about 20 similar villages if the age-distributions had been similar, and a larger number as the age distributions were not similar.

The observed distribution pattern for skin cancer (no skin cancer cases for exposure level below 150 ppb) suggested a threshold effect of arsenic exposure on skin cancer. To explore such a potential threshold effect, the prevalences of the four skin disorders were fitted as functions of arsenic exposures, using a simple linear regression model and using the x-intercept of the least squares regression line as an estimate of the threshold value. The least squares regressions for both PAC and CAD exposures were statistically significant for all four skin disorder groups. For the PAC exposures, the predicted threshold levels (Table 5.2) were 43 ppb for skin cancer, 4.9 ppb for keratoses, 1.4 ppb for dyspigmentation, and 14 ppb for combined keratoses and dyspigmentation. Of the four estimated threshold values, only the one for skin cancer was significantly different from zero at p = 0.05 level. Similar patterns of x-intercepts were also found with CAD exposures (Table 6.2). Skin cancer had the largest x-intercept (313 ppb-years; p = 0.12), approximately an order of magnitude greater than for keratoses (43 ppb-years; p = 0.24), for dyspigmentation (35 ppb-years; p = 0.65), or for both combined (25 ppb-years; p = 0.68). None of the four outcomes had an x-intercept significantly different from zero with CAD exposures in the simple linear model analysis.

The prevalences of the four skin disorders were also fitted as functions of arsenic exposures, using the hockey-stick model by a maximum likelihood method. The regression was statistically significant for all four of the skin disorders and for both types of exposure measure. The predicted threshold levels (d_t) for the PAC measure (Table 5.3) were 122 ppb for skin cancer, 42 ppb and 30 ppb for keratoses, 50 ppb and 47 ppb for dyspigmentation, and 42 ppb for combined keratoses and dyspigmentation. The threshold level for skin cancer was about three-times higher than that for the non-malignant skin disorders. All four of the outcomes had an x-intercept that was significantly different from zero with PAC exposures in the hockey-stick model analysis.

The threshold effect was also examined for CAD exposures (Table 6.3). While the threshold level for skin cancer at 617 ppb-yrs was higher than those of the non-malignant conditions, the skin cancer threshold level was statistically significant for one of three acceptable fits (617 ppb-yrs) at the P <0.05 (one sided) level.. The predicted threshold values for the non-malignant skin conditions were about 400 ppb-yrs with values 353 ppb-years and 135 ppb-yrs for keratoses, 440 ppb-yrs for dyspigmentation, and 406 ppb-yrs for both combined. Keratoses, dyspigmentation, and the two non-malignant conditions combined all had x-intercepts that were significantly different from zero at p <0.05 with CAD exposures in the hockey-stick model analysis. For dyspigmentation alone, there was also an acceptable but considerably less probable fit where the confidence limits included zero. The malignant outcome had an x-intercept that was significantly different from zero with a one sided fit at p <0.0.05 for the most probable fit with CAD exposures in the hockey-stick model analysis.

The interpretations of the three analytic models are similar. The frequency-weighted model analysis shows that skin cancer risk is non-linear with respect to the arsenic exposure level (i.e., observed number of cases with exposure < 150 ppb was significantly fewer than predicted by the linear model). This observation is consistent with the thresholds indicated by the least squares linear models and the hockey-stick models. The least squares linear model and the hockey-stick model are both maximum likelihood estimate models. The hockey-stick model has an advantage over the least squares linear model in that it is sensitive to both the numerator and the denominator of each prevalence point and thus considers the weight of evidence at each point. The least squares linear model is an analysis of the set of prevalence points, without consideration of the sample size of each prevalence point. In addition, it is somewhat unrealistic in permitting solutions with P < 0 or > 1 in some situations. Thus, the hockey-stick model is more likely to approximate the central tendency of the underlying data.

The analytic results with this dataset from Inner Mongolia in the 1990s are remarkably similar to those from southwest Taiwan in the 1960s. Tseng et al. published their skin cancer prevalence data with well-water arsenic exposures in 1968.⁴ Analysis of those data (Table 7.2), with the weighted means of the exposure intervals used by Byrd et al. yielded an x-intercept of 118 ppb by the simple linear model and 119 ppb by the hockey stick model (Figure 8.2).⁵ EPA’s risk assessment for the Taiwan skin cancer prevalence in Taiwan was based on a non-threshold model that did not permit the examination for a threshold. The exposure strata for the Taiwan analysis and the Inner Mongolia analysis differ. Nonetheless, their findings can be roughly compared. Both for Taiwan at exposures < 300 ppb and for Inner Mongolia at exposures < 500 ppb, the skin cancer prevalence is 0.2 %. The skin prevalence rate in the two highest exposure strata in the Inner Mongolia data (150 +) is 1.75 %, and the skin cancer prevalence rate in the two highest exposure strata in the Taiwan date (300 +) is 1.79 %. Thus, the dose-related skin cancer prevalence rates in the Inner Mongolia and Taiwan data sets are roughly equivalent.

Table 8.2. Frequency and prevalence of skin cancer by weighted mean arsenic concentration intervals (ppb) in the Taiwan study (Tseng et al., 1968)

As Concentration (ppb)                                     Skin Cancer                 Population                               Prevalence
Range            Weighted Mean                                  (N)                            (N)                                          (%)
< 300                       171                                        21                             9,526                                        0.2
300-600                   473                                        60                             5,413                                        1.1
> 600                        785                                     185                             8,251                                        2.2
Total                         460                                      266                          23,190                                         1.1

8.6. Time considerations of PAC Exposures

Time as a factor can be expressed as age, as duration of exposure, and as latency (time since exposure). Analyses have been conducted to examine each of these with respect to both the arsenic exposures and the outcome measures.

Section 7 presented analyses that accounted for age distribution differences in the exposure strata by age-adjusting the prevalence data. Table 7.1 and Figure 7.1 presented the analyses for the PAC exposure strata. Section 8.8 below presents a formal latency analysis with results shown in Table 8.5. Both mean age in years and mean duration of time since peak exposure in years are presented as time variables in both tabular analysis in Table 8.3 and in graphic analysis in Figure 8.3.

Table 8.3 Time Variables and Clinical Prevalences by Peak Arsenic Conc. (PAC) strata

                                       Time Variables                  Exposure               Clinical Outcome  Prevalence (%)
PAC       Subjects             Age Duration                     Dose.                  Kerato      Dyspig        Kerato+      Skin
                N (Mean)                  from Peak*             (Arith-mean)                                               Dyspig     Cancer

<10        287                    26.0       21                            5 0.                        3             1.0               0.3         0.0
10-         405                    28.6      17.5                        15 0.                        2              0.0              0.0         0.0
30-         412                    29.6      21.5                          33                        1.0            1.0               0.7         0.0
50-         516                    32.5      20.3                          55                        1.6             1.4              1.2         0.0
60-         565                    37.6      22.0                          70                        4.1             1.9              1.9         0.0
100-       536                    34.3      18.7                         122                     11.4             7.8              5.8         0.0
150-       416                    35.7       20.2                        175                     10.8             8.2              5.8         1.2
500+        42                    37.1       15.5                      1,048                    69.0            47.6            42.9         7.1

* Time interval in years from beginning of peak exposure until examination date (1992).

These analyses are presented simultaneously for the four clinical conditions.Table 8.3 presents the PAC strata with the mean age, mean exposure duration since beginning of peak exposure (time interval to 1992, the date of examination), mean peak exposure, and the prevalence rates for each of the four clinical outcomes.

No new strong patterns stand out in a visual inspection of Table 8.3. The mean age of the subjects seems to be higher for those with greater PAC levels. This has been adjusted for in the age-adjusted analyses shown in section 7 and are presented here without age-adjustment. How strongly an age trend is seen is dependent upon how the strata are grouped. The mean age for those with PAC < 30 is 27.2 years, for those with PAC between 30 and 60 is 31.8 years, for those with PAC between 60 and 150 is 35.9, and for those with PAC of 150 ppb or greater is 35.8 years.

Although there appears to be a positive age-related trend, the latency appears to have a negative, if any, effect. The mean time since peak exposure until examination appears to be about 20 years in all PAC strata. There appears to be no relationship between the level of the peak exposure and the duration of time between the initiation of the peak exposure and the date of examination (Duration from Peak).

A graphic analysis of these data has been developed (Figure 8.3) which reveals the pattern of association over the PAC range of the time variables, the exposure variable, and the clinical outcome prevalences. Because of the wide numeric range of values seen in Table 8.3, each column of values have been normalized to the maximum value in the column and expressed as a percentage of that value. Figure 8.3 shows the distribution of those relative values across the PAC strata for the time values (mean age and duration from peak), exposure value (arithmetic mean of the arsenic exposure [ppb]), and the clinical values (prevalence of the four clinical conditions).

Figure 8.3 demonstrates that the time variables show little variation across the PAC exposure strata and maximize in the 30-60 (+) ppb range. In contrast, the exposure variable (Arith-mean) and the non-cancer clinical variables show a gradual rise (similar to the exposure) and maximize in the final strata. The cancer variable first appears in the 150 (+) ppb strata and maximizes prevalence in the 500 (+) ppb strata.

Similar analyses can be developed assessing time considerations in this data set using the cumulative arsenic dosage (CAD) as the measure of arsenic exposure. In this case, time as a factor is expressed as age and as mean exposure years (duration of well-use history). Table 8.4 presents the CAD strata with these two measures of time consideration and the prevalence rates for each of the four clinical outcomes. The cumulative arsenic dosage is related both to the mean age of each group and to the number of years of exposure history reported in the data set. The critical question, however, is what exposure variables are related to the prevalence rates of the clinical conditions. Figure 8.4 presents the data in Table 8.4 in the same manner as Figure 8.3 presented the data in Table 8.3 – i.e., each column of values has been normalized to the maximum value in that column and expressed as a percentage of that value. Figure 8.4 demonstrates that the mean cumulative arsenic exposure, but not the time variables, show the same patternacross the exposure strata as do the relative prevalences of the clinical conditions. Skin cancer is only seen in the top two exposure strata – i.e., at 3162 ppb-years to 10000 ppb-years and above.

Table 8.4 Time Variables and Clinical Prevalences by Cumulative Arsenic Dosage (CAD) strata
 

CAD	Mean Age	Mean Years of Exposure Reported	Mean CAD (ppb-yr)	Keratoses	Dyspig	K+D	Skin Cancer
<10	8.6	1.0	5	0.0	0.0	0.0	0.0
10-	13	3.4	20	0.0	0.0	0.0	0.0
32-	14	6.7	62	0.0	0.6	0.0	0.0
100-	22	15	203	0.04	0.4	0.2	0.0
316-	27	20	617	1.5	0.9	0.6	0.0
1000-	39	31	1893	6.9	5.0	3.9	0.0
3162-	53	46	4656	17	11	9.2	0.4
10000	53	48	12971	46	32	24	0.8

The associations between time variables, exposure, and clinical outcome prevalences are shown similarly in the CAD analysis as in the PAC analysis. However, here both mean age and mean years of exposure increase as mean CAD exposure increases, which is expected. The clinical prevalences also do not differ from what is expected.

Much information on latency is suggested from preliminary review of the data. Each strata in the peak arsenic concentration (PAC) analysis has had an average of 15 to 22 years since the initiation of their peak exposure (Table 8.3). Thus, it is quite likely that a "sufficient" latency period exists within this data set for clinical conditions attributable to those peak exposures to be observed. With respect to skin cancer, six of the eight cases had their peak exposure more than forty years prior to the dermatological examination.

Similarly, the cumulative arsenic dosage (CAD) analysis (Table 8.5) reveals that for those with 100 ppb-year or greater cumulative arsenic dosages, the mean duration of observation increases from 15 years to over 45 years. This period of observation should provide for considerable latency consideration for each of the clinical conditions observed. The group of subjects with cumulative arsenic exposures of less than 100 ppb-years may not have had sufficient duration of observation for certain of the clinical conditions to be attributable; however, they have shown little evidence of disease and do not represent a great proportion (< 10 %) of the study population.

Previously, a case-control analysis of the data from Tie Men Geng and Zhi Ji Liang showed that the chronic arsenicism cases were statistically significantly more likely to have had at least ten years of exposure (Odds Ratio = 4.0; 95 % confidence limits of 1.6 and 9.9) and to have a peak exposure of greater than 200 ppb (Odds Ratio = 13.3, 95 % confidence limits of 5.4-32.8).⁵ Cases and controls met the same minimum exposure criteria and were matched for gender, age, occupation, education, and living and working conditions.

A formal latency analysis of these data has been undertaken in which the maximum likelihood hockey-stick model was examined using latencies of 0, 10, and 25 years. In these analyses, only the exposure occurring more than 0, 10, or 25 years before the time of observation were taken into consideration. The PAC and CAD data were separately analyzed for each of the clinical conditions. Each data set was fit to an ?, ? parameter model with a determination of ?² and p-value. The threshold (d_t) for each good fit was identified. The 95% confidence limits of the threshold was calculated in the following way: while all other parameters were fixed, the fitted threshold value was increased and decreased until the ?² changed by 2 units, corresponding to 2 standard deviations of the parameter, or 95% confidence intervals. If zero was excluded from the 95% confidence limit, the fitted threshold was considered to be statistically significant. Table 8.5 presents the statistically significant fitted thresholds from models with good fit to the data (p > 0.05) for each clinical condition and using either the PAC or CAD exposure data.

Table 8.5 Statistically significant threshold values for latency models with good fit using either PAC or CAD exposures and latencies of 0, 10, and 25 years for each clinical condition.

Clinical Condition                                                                        PAC (ppb)                                              CAD (ppb-yrs)
              (Latency)                                                                0 yrs      10 yrs      25 yrs                 0 yrs           10 yrs        25 yrs
Keratoses                                                     29, 42           -                 -               135, 353             -            -
                                                                               (Mean = 36)                                                    (Geomean = 218) 
Dyspigmentation                                           47, 50           -                 -                  440                   -           -
                                                                (Mean = 49)
Keratoses                                                       42              -                 -                   405                   -           -
with Dyspigmentation
Skin Cancer                                                122 168,     299     167, 312                    -               5771 1277, 2800
                                                                                      (Mean = 234)                    (Mean = 240)                                                         (Geomean = 1890)

Significant threshold values for good fitting models were found for all conditions under zero latency conditions and only for skin cancer under either 10 year or 25 year latency conditions for the PAC and the CAD exposure analyses. For non-malignant arsenic skin disorders using the PAC exposures, the threshold values were between 29 and 50 ppb with a mean at 42 ppb. For non-malignant arsenic skin disorders using the CAD exposures, the threshold values ranged between 135 and 440 ppb-years with a mean of 333 or 465 ppb-years. Only skin cancers showed significant thresholds in good fitting models under 10- or 25-year latency conditions in either the PAC or CAD exposure analyses. For skin cancer using PAC exposures, the threshold values ranged between 167 and 312 ppb for the 10- and the 25-year latency conditions with a mean of 236 ppb. For skin cancer using CAD exposures, the threshold values was 5771 ppb-years under 10 year latency analysis and was 1277-2800 (geometric mean = 1890) ppb-years under 25 year latency conditions.

The lowest significant threshold in a good fitting model for skin cancer in the PAC exposure analysis was at 122 ppb under the zero latency condition. Under the 10 and 25 year latency conditions, the significant threshold in good fitting models for skin cancer in the PAC exposure analysis ranged from 168 to 312 ppb with a mean of 236 ppb.

In the zero-year latency analysis of the PAC exposure data, only the 122 ppb threshold for skin cancer came from a model with an excellent fit (goodness-of-fit p > 0.90). The p-value for the goodness-of-fit of the combined clinical finding of keratoses with dyspigmentation was 0.33. The p-values for the models from which the remaining thresholds in Table 8.5 for zero-year latency analysis of the PAC exposure data were derived were in the 0.05-0.15 range. The p-values of the PAC exposure models with 10-year latency were 0.57 for the 168 ppb threshold and 0.15 for the 299 ppb threshold. The p-values of the PAC exposure models with 25-year latency were 0.16 for the 167 ppb threshold and 0.08 for the 312 ppb threshold.

In the zero-year latency analysis of the CAD exposure data, all the thresholds shown in Table 8.5 came from models with an excellent fit (goodness-of-fit p > 0.90). In the ten-year latency analysis and the 25-year latency analysis of the CAD exposure data, the statistically significant thresholds all came from models with goodness-of-fit p-values in the 0.45-0.75 range.

The latency analyses of the PAC exposure data show for the non-malignant arsenical skin findings that thresholds are only identifiable in the zero-year latency analysis and are not identifiable in the ten-year and 25-year latency analyses. In the PAC exposure data analysis for the malignant arsenical skin finding (skin cancer), the only statistically significant threshold that comes from an excellently fitting model is the 122 ppb threshold in the zero-year latency analysis. In the latency analyses for 10 years and 25 years, the statistically significant threshold values remain in the 150-300 ppb range and the good-fitting models have lower p-values.

The latency analyses of the CAD exposure data show for the non-malignant arsenical skin findings that statistically significant thresholds are identifiable in the zero-year latency analysis but not in the ten-year and 25-year latency analyses. These identified thresholds for zero latency are all under 500 ppb-years and come from fits that are excellent. For latencies > 10 years and > 25 years, the goodness-of-fit was significantly worse (p ~0.03 –0.05) and the thresholds were not statistically significant. This suggests that the latency for the non-malignant skin findings is likely to be less than ten years. Thresholds for skin cancer continue to be identified only in the ten-year and 25-year latency analyses, suggesting a latency 10 years or greater. Six of the eight skin cancer cases had exposure histories that exceeded forty years.

There has been some suggestion that concentrations recently measured in the wells are lower than in the past when the major exposures occurred and that exposures preceding the beginning of the well-use histories may have led to an under ascertainment of exposure. Such misclassification of exposures would have suppressed the appearance of a threshold, as more cases would be classified as being below the apparent threshold than actually occurred. Thus, such misclassification of exposures or of their associated skin lesions would spuriously decrease (not increase) the evidence for a threshold.

The Inner Mongolia data demonstrate that both hyperkeratoses and dyspigmentation are observed at lower arsenic exposure levels than skin cancer. The Inner Mongolia study found eight cases of skin cancer among 3,179 arsenic-exposed well users. All eight cases had both hyperkeratoses and dyspigmentation, and no cases were observed among the 3,085 subjects who did not have both hyperkeratoses and dyspigmentation. Further, only eight of the 94 subjects with both hyperkeratoses and dyspigmentation developed skin cancer. These observations indicate that hyperkeratoses and dyspigmentation are not sufficient pre-conditions for arsenic-induced skin cancer but do suggest that they may be necessary pre-conditions. Both the Inner Mongolia and the Xinjiang studies have demonstrated a greater prevalence of hyperkeratoses than of dyspigmentation, while the Taiwan and Bengal studies have demonstrated a greater prevalence of dyspigmentation than of hyperkeratoses.^2,4,7 All studies have shown a frequent co-prevalence of both hyperkeratosis and dyspigmentation, and all studies show a far higher prevalence of hyperkeratoses and dyspigmentation than of skin cancer.

The exposure-response curves in these analyses reveal that the prevalence of keratoses, dyspigmentation, and skin cancer strongly depend on the level of arsenic exposure. A strong increase of response with increase in dose exists whether the peak arsenic concentration (PAC) or the cumulative arsenic dosage (CAD) was used as the exposure measure. The dose-response relationship was apparent for all four of the skin disorders examined, though the pattern of the dose-response relationship tended to vary among the different skin disorders. A threshold below which no arsenic skin lesions are observed seems likely but not certain.

With respect to the peak arsenic concentrations (PAC) in the drinking water, both the non-malignant and the skin cancer prevalence appeared to follow a threshold model. The data for the non-malignant lesions seemed definite in showing a threshold in the vicinity of 40-70 ppb. For the malignant skin lesions, the absence of events at low concentrations strongly suggests a threshold higher than that for non-malignant lesions although the paucity of events prevents firm conclusions.

With respect to the cumulative arsenic dosage (CAD) in the drinking water, the skin cancer prevalence tendency to follow a threshold model is only seen in the latency models. The evidence for a threshold effect on non-malignant skin disorders is only present in the zero latency model and appears to be at a lower threshold level than might be foundfor skin cancer. The fits to a hockey stick function, adjusted to limit the prevalence to 100%, fitted all lesions well. Although a threshold was suggested for all lesions, this was only statistically significant for keratoses (about 240 ppb-yrs) and keratoses plus dyspigmentation.

Analyses of time considerations suggested that arsenic exposure level, but not age, exposure duration, or time since exposure, was greatly related to the prevalence of the skin disorders in the PAC analysis. Furthermore, sufficient latency has occurred between time of exposure and time of observation for the clinical outcomes to be assessed as attributable to the exposures. The latency analyses indicate that the latencies for the non-malignant skin effects may be less than 10 years and that the latencies for skin cancer may extend into the 10-25 year range. The data are suggestive that the presence of hyperkeratoses and dyspigmentation may be necessary, but not sufficient, precursors for arsenic skin cancer; but that is not proven.

The PAC analysis predicts that the study population is made up of two sub-groups, those that have had exposure above the threshold level and those who have not. The frequency-weighted PAC exposure analysis showed a cut-point at 150 ppb arsenic, a level we shall use in subsequent analysis. It may be that the distribution of cumulative arsenic exposures reflects the distribution of the nearly 3,000 residents who are recorded as never having been a user of a well with an arsenic concentration of > 150 ppb, rather than the distribution of the less than 500 who acquired risk by being a user of such a well (or wells).

Table 9.1 presents the CAD exposure strata distribution of the 458 residents who had used the > 150 ppb wells. Application of the average cumulative exposure in that strata to the number of persons in that strata yield a calculation of the number of cumulative arsenic exposure (in ppb-yrs) experience by those individuals. An analytic question that can then be examined is whether the observed distribution of skin cancer cases in these strata follows that which would be predicted based on the distribution of the number of ppb-yrs in each strata (i.e., is directly dependent on the cumulative arsenic exposure) or that which would be predicted based on the distribution of the number of people in each strata (i.e., is independent of the cumulative arsenic exposure). The chi-square analyses in Table 9.1 demonstrate for both the assumption that the risk in the high exposure group is independent of the cumulative exposure and the assumption that the risk in the high exposure group is dependent upon the cumulative exposure. The critical chi-square value for a two-tailed p-value < 0.50 with three degrees of freedom is 7.82. In neither case is the chi-square value greater than the appropriate critical value. At that level, one would be able to state that the assumption that predicted a case distribution significantly different from the observed distribution could not explain that observed case distribution.

Table 9.1 For subjects with PAC > 150 ppb and by CAD strata, the Observed Number of Skin Cancers and Numbers of Skin Cancers Predicting Assuming Dependence and Independence on Cumulative Arsenic Exposure (ppb-yrs).
 

CAD	N (obs)	Ave ppb-yrs	Total ppb-yrs	n (obs)	n (Independ)	Chi-Sq	n (Depend)	Chi-Sq
316-	44	617	27,148	0	0.77	0.77	0.13	0.13
1000-	207	1893	391,851	4	3.62	0.04	1.85	2.48
3162-	170	4656	791,520	3	2.97	0.00	3.75	0.15
10000-	37	12971	479,927	1	0.65	0.19	2.27	0.71
	458		1,690,446	8	8.00	1.00	8.00	3.47
						dF = 3		dF = 3

In this circumstance, both chi-squares are well below the critical value, which indicates that these data are insufficient to distinguish between the two opposite assumptions. Figure 9.1 shows that the observed distribution of cases is quite similar to the distribution predicted by the assumption of independence and that the distribution predicted by the assumption of risk dependent on the cumulative arsenic exposure is displaced to the right and away from the observed distribution. This visual presentation reaches the same conclusion – that the data are insufficient to distinguish between the two assumptions. Nonetheless, they do suggest (1) that a larger study may be able to distinguish and (2) that the appropriate area for study of skin cancer risk with arsenic ingestion is in the moderate arsenic exposure (100 ppb-1000ppb) rather than either in the low exposure range of < 50 ppb or the high exposure range of > 1000 ppb.

This report presents the dose-response data analysis for skin cancer and other dermatological effects of arsenic exposure among the population of three villages in Huhhot, Inner Mongolia with arsenic-contaminated wells. Each subject was examined by physicians from the local health department, and dermatological findings of hyperkeratoses, dyspigmentation, and/or skin cancer were recorded. The well-use histories for these subjects were obtained, as were the arsenic measurements from these wells. The mean arsenic level was used to represent the arsenic level of wells that had more than one measurement. Based on the well-use histories and laboratory values of the arsenic content of the well waters, two measures of arsenic exposure were developed for each subject. The first measure was the peak or highest arsenic concentration (PAC) for the individual, based on their well-use history. The second measure was the cumulative arsenic dosage (CAD) that was the summation of the exposures and durations for each well used, summarized as ppb-years. Exposure strata were developed, and the prevalences of dermatological findings (particularly skin cancer) across the exposure strata were examined.

Four measures of dermatological disorder were analyzed – (1) hyperkeratoses, (2) dyspigmentation, (3) hyperkeratoses with dyspigmentation, and (4) skin cancer. Analyses of the distribution of the prevalence of specific dermatological disorders across exposure strata were conducted. Analyses were conducted using a frequency-weighted model as well as both a simple linear model and a hockey-stick model and later with a formal latency analysis.

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.

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. Analysis with respect to the cumulative arsenic dosage (CAD) is consistent with the analysis of the peak arsenic concentration, but less clear. No skin cancer was observed among those whose peak arsenic concentration was less than 150 ppb or whose cumulative arsenic dosage was less than 1000 ppb-years. 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.

Additional analyses could be considered, but the power of this study to further describe the dose-response relationship between arsenic ingestion and skin cancer is limited by the identification of only eight cases of skin cancer in this population at the time of their examination. Observations made from the analyses of these data should be used in the design of further studies. These observations should guide the selection of the study population by exposure history. Subsequent study of this population ten years after the initial study, or extension of this study to a larger Inner Mongolian population may be considered. The design of subsequent studies should be dependent upon the questions whose answers are sought. The evidence presented here of a threshold arsenic exposure level for skin cancer that is consistent with the evidence from the Taiwan study of Tseng should be examined in other populations. Such evidences should be taken into consideration in attempting to establish safe drinking water standards for arsenic exposure.

(1)  Luo ZD, Zhang YM, Ma L, Zhang GY, He X, Wilson R, Byrd DM, Griffiths JG, Lai S, He L, Grumski K, and Lamm SH. (1997) Chronic arsenicism and cancer in Inner Mongolia – consequences of well-water arsenic levels greater than 50 ug/l. In Arsenic: Exposure and Health Effects. Edited by Abernathy CO, Calderon RL, and Chappell WR, Chapman and Hall Press, London Chapter 5, pages 55-68.

(2) Niu S, Cao S, and Shen E. (1997) The status of arsenic poisoning in China. In Arsenic: Exposure and Health Effects. Edited by Abernathy CO, Calderon RL, and Chappell WR., Chapman and Hall Press, London Chapter 7, pages 78-83.

(3) Cebrian ME, Albores A, Aguilar M, and Blakely E. (1983) Chronic arsenic poisoning in the North of Mexico. Human Toxicol. 2:121-133.
 

(4) Tseng WP, Chu HM, How SW, Fong M, Lin CS, and Yeh SHU. (1968) Prevalence of skin cancer in an endemic area of chronic arsenicism in Taiwan. Journal of the National Cancer Institute. 40:453-463.

(5) Byrd DM, Roegner ML, Griffiths JC, Lamm SH, Grumski KS, Wilson R, and Lai S. (1996) Carcinogenic risks of inorganic arsenic in perspective. Int Arch Occup Environ Health. 68:484-496.

(6) Byrd DM, Sadje R, Lai S, Luo ZD, Zhang YM, Ma L, Zhang GY, He XZ, Lamm SH, Grumski K, and Wilson R. Risk of skin lesions in relation to arsenic exposure. Presented at the Second International Conference on Arsenic and Health, San Diego, California, June 1995.

(7) Mazumder DNG, Chakraborty AK, Ghose A, Gupta JD, Chakraborty DP, Dey SB, and Chattopadhyay N. (1988) Chronic arsenic toxicity from drinking tubewell water in rural West Bengal. Bulletin of the World Health Organization. 66:499-506.

(8) Fan Chengwan, Naren Gaowa, Zhang Yumin, et al. (1993) Analysis for arsenical water and approach for reason of rich arsenic in Western Huhhot Basin. Environment and Health (Beijing), 10(2):56-58. [in Chinese]

(9) Luo Zhendong, Zhang Yumin, Ma Liang, et al. (1993) Epidemiological survey of chronic arsenic poisoning at Tie Mengeng and Zhi Jiliang villages in Inner Mongolia. Chinese Public Health (Beijing), 9(8):347-348. [in Chinese]

(10) Zhang Yumin, Ma Liang, Luo Zhengdong, et al. (1994) Water quality analysis of arsenic-enriched groundwater in the large area of Western Huhhot Basin. Rural Eso-Environment, 10(1):59-61. [in Chinese]