IntroductionMany apparent associations later turn out to be false. When does a study show causation between a risk factor or treatment and the disease? Even if an association is established between a risk factor and a disease it does not necessarily follow that the risk factor caused the disease. Maybe a third factor caused both or perhaps early forms of the disease result in the so-called risk factor so that causation works in the opposite direction. For example:
- Third Factor. In a Jan 2009 paper, researchers found that prostate cancer patients had more frequent sexual activity when young than non-prostate cancer patients. The authors did not conclude that sexual activity in youth causes prostate cancer, despite the association. In particular, they did not conclude that modifying one's sexual activity would modify the risk for prostate cancer. Rather, they hypothesized that there was a third factor, hormone levels, that caused both. [news item] [PMID: 19016689 If their hypothesis were correct this would be an example of association without one of the associated factors causing the other (because a third caused both).
- Reverse Causality. In a 2007 nomogram (see Figure 3a of [Full Text] [PMID: 17513807] and also this online calculator) to predict recurrence after radiation therapy it appears that the risk declines with increasing dosage until dosage reaches about 66 Gray and then, oddly, the risk increases again from that point onward. This was based on retrospective data and one wonders whether this is reverse causality at work. That is, perhaps the patients who received the highest dosage of radiation did so because they were known to be at higher risk rather than being at higher risk because they got the higher dosage. Fat people are the ones that drink diet coke (association) but that does not mean that diet coke makes people fat (causation). Sometimes the direction of causality, if it exists at all, seems unclear. A study of 34,000 adults found that those who were the most satisfied with their health care had a 26% higher mortality rate than those least satisfied. [Health News Review] [PMID: 22331982] Its hard to know if there is causality here and which direction it acts if it is present. For example, perhaps dissatisfaction leads patients to better pursue appropriate treatment or perhaps those who are more satisfied tend to be over-treated and the over-treatment results in poorer health outcomes. To dig deeper we would need to know more.
ConfusionConfusion surrounding association vs. causation is everyhwere:
Media Confusion. The confusion between association and causation can often be seen in media reports. A headline to a news article Excess Weight Increases Prostate Cancer Mortality makes such an error. The report itself (and even the text of the news item) make no claim of causation but the headline, which may have been written by a different person, jumps to the unwarranted conclusion of causation. The investigators correctly avoided assuming that simply because excess weight and prostate cancer were associated that prostate cancer was caused by excess weight. For example, if any of these were true then we might see this association even in the absence of such causality: (1) obese men delay longer in seeking treatment, (2) diagnostic procedures are less accurate for obese men so their condition is missed more often and they do not receive timely treatment -- in fact it is known that obesity reduces PSA levels since the obese have a greater volume of blood diluting the PSA in the blood further thereby making it less likely that cancer will be detected in the obese [Full Text] [Journal Watch comment], (3) genes which predispose one to prostate cancer also predispose one to obesity or (4) men with prostate cancer respond by over-eating (i.e. causality works in reverse).
The confusion in media can be readily seen in the reporting of two large screening studies for prostate cancer, the Prostate Lung Colorectal Ovarian Cancer (PLCO) and the European Randomized Study of Screening for Prostate Cancer (ERSPC). As reported in [PMID: 21446935] 23% of the newspaper reports concluded that screening is positive and 31% concluded that screening is negative. 78% of UK media concluded that the studies showed that there was insufficient screening whereas 57% of US media and 80% of Canadian media concluded that screening is excessive. Clearly one has to be very careful in reading such reports since there was no consistent interpretation of the result. The web site Health News Review reviews media health reports and from reading some of those reviews one can get a very quick idea of the poor standard in medical reporting. This shows that it is important to be able to assess studies independently of the media reports.
Patient Confusion. Patients themselves may confuse co-occurrence with causation. Assessing causation is much more difficult than it may appear on the surface and patients may not be able to evaluate it. The result is mistaking coincidence or other reasonable explanations for miraculous cures. For example, in this March 9/08 post by a patient who claims to have been cured of prostate cancer by taking cayenne pepper [link] he more likely simply had a false negative biopsy. Another patient claims that simply not drinking milk will cure prostate cancer and points to (1) his reduction of PSA after eliminating milk and (2) increase after resuming it. But he was also receiving conventional medical treatment which could easily account for the reduction in PSA and subsequent recurrence as could chance and numerous other factors. Both these relate to items that do have biological plausibility to them, which is one of the Bradford Hill criteria, and it may even be that the measures which appeared to work in these isolated cases turn out to be highly effective but the point is that these isolated examples should not be confused with "proof". (For more on cayenne pepper see [PMID: 16540674] and for dairy and calcium see the page labelled 129 and following of the WCRF report.)
Study Confusion Even published studies themselves may have limitations, biases, errors or other problems. Sometimes the investigators are well aware of these but the intended audience is relatively sophisticated and would be expected to interpret the results correctly whereas in other cases they are oversights or limitations in the study. See the three part post that starts here for a fuller discussion of why studies can be wrong. We also discuss some examples below.
Simpson's Paradox. In investigating possibly discriminatory behavior in prostate cancer treatment it was found that at University Hospital that: (1) nearly the same percentage of white men and black men were offered prostate cancer surgery (59% white, 58% black). (2) At VA Hospital the percentages were also the same (28% white, 28% black). Yet overall 50% of white men were offered prostate cancer surgery and only 32% of black men. How is it possible that black men were offered surgery less often in total even though they were offered it just as often as white men at each hospital. The actual numbers from the study were:
|University Hospital||54/91 (59%)||7/12 (58%)||61/103 (59%)|
|VA Hospital||11/40 (28%)||22/79 (28%)||33/119 (28%)|
|Overall||65/131 (50%)||29/91 (32%)|
The CriteriaA landmark paper which is one of the most cited works in health research is the 1965 address of Sir Austin Bradford Hill in which he discussed 9 factors (the Bradford Hill Criteria) that support casuation:
- Strength. Its hard to explain away the fact that the death rate from lung cancer in smokers is 9-10 times that of nonsmokers. Some other factor would have to have a very large effect to overcome that making it much less likely that its something else. In fact there is a mathematical condition known as Cornfield's condition that states that the apparent effect cannot be reversed by the confounder if the effect of the confounding factor is not greater than the effect of the apparent factor. (In the table above where the apparent factor was race and the confounding factor was hospital the ratio of hospital (row) totals is (59% / 28% = 2.1) which exceeds the ratio of race (column) totals (64% / 47% = 1.4) leading to the possibility of reversal in conclusion. [article on Cornfield's condition].
- Consistency. If many studies come to the same conclusion its more powerful than if that were not the case. The cause of inconsistent studies can sometimes be traced to an unaccounted for and possibly unknown factor whose presence affects the efficacy of the treatment. For example, Javier A. Menendez at the Catalan Institute of Oncology in Girona, Spain, and colleagues found that extra virgin olive oil was effective against breast cancer cells in HER2 positive individuals but not in HER2 negatives. At the time of the trials it was likely not known what the HER2 status of the subjects was but with this information new trials could be designed that stratified on HER2 analysing HER2 positive and negative subjects separately possibly showing effect by focusing on the subgroup, HER2 positives, where the effect is expected without dilution from the HER2 negative group where no effect is expected. See [WebMD article] [PMID: 19094209] [NY Times, Dec 29/08].
- Specificity. Hill summarized this as "one cause, one effect". If a particular risk factor raises the death rate of a specific type of disease and not just the general death rate then its clearly easier to trace the cause to the disease than if it were more diffuse. In general, we wish to isolate the cause and the effect. Sometimes it can be difficult to distinguish which cause is the real one. For example, it might appear that prostate cancer patients with diabetes are at higher risk of mortality (from prostate cancer or other cause); however, a paper at 2008 ASCO found that that was because diabetics tend to be overweight. When they controlled for weight, e.g. compare diabetics and non-diabetics who are normal weight and compare diabetics and non-diabetics who are overweight, they found that diabetics had the same risk as non-diabetics. Being overweight was the true risk factor.
- Temporality. Does the cause come before the effect? If the effect came before the cause one might question whether it really is the cause. Example 1. For example, if we are looking to assess whether Agent Orange causes prostate cancer we would want to ensure that the exposure to Agent Orange preceded the prostate cancer. Example 2. One patient remarked that those with more aggressive prostate cancer seemed weak but were they weak before they had it or did the cancer or treatment make them weak (reverse causality)? Which one came first would be essential to know. Example 3. Nobel Prize winning biochemist Otto Warburg noted that cancer cells generate energy using glycolysis whereas normal cells use oxidative phosphorylation and it had been hypothesized that this was an adaptive response to the oxygen deprived conditions in the tumor; however, evidence since then has determined that glycolysis emerges prior to the tumor's exposure to hypoxic conditions so the adaptation hypothesis must be rejected and we must look elsewhere for an explanation. A new theory has since emerged that adpating to the uptake of nutrients rather than to energy production is the driving force. [PMID: 19460998 [Full Text] This theory is currently supported by complex computer models which show that glycolysis is implied by enzyme constraints whereas it fails to emerge in the absence of those constraints. [PMID: 21423717] Example 4. A study in another field of health concluded that children given antibiotics had greater asthma but that ignores that they were given the antibiotics because of some previous condition and such conditions include lung infections. Example 5. Dietary studies are particularly problematic for prostate cancer since it is generally impossible to establish that the diet preceeded the onset of prostate cancer given the long amount of time that it takes for the disease to develop.
- Dose Response. If consuming greater amounts of a substance gives more protection and lesser amounts less protection that is greater evidence of benefit than if the relationship does not vary in an increasing way. A similar argument can sometimes be made even when "dose" is something other than medication: "Inside medical institutions, in counties containing teaching hospitals, fatal medication errors spiked by 10% in July and in no other month [JR = 1.10 (1.06-1.14)]. In contrast, there was no July spike in counties without teaching hospitals. The greater the concentration of teaching hospitals in a region, the greater the July spike (r = .80; P = .005). These findings held only for medication errors, not for other causes of death." [PMID: 20512532]. Here "dosage" is the concentration of teaching hospitals and response is the number of fatal medication errors. The fact that the response increases with "dosage" here provides greater evidence of a causal effect than might otherwise be implied.
- Biological Plausibility. If there is plausible biological mechanism that could explain the relationship the possibility of causation is increased.
- Coherence. The possibility of causation should not contradict other known facts. For example, if people with higher intakes of selenium have less prostate cancer then we might question whether the selenium caused this if we found that their blood serum levels of selenium were not also higher. For an example from the literature, consider this May 2008 study [PMID: 18467718] aiming to compare laparascopic/robotic and open prostatectomy outcomes. However, as pointed out by Michael Blute in the same issue of the journal in which the study was published [PMID: 18467714] [Full Text]: "the data in this study indicate that the complication rates to open surgery may not be consistent with that reported for optimal open surgery."
- Experimental Evidence. Well designed experiments may give strong reason to believe that causation is at work. (a) Randomized clinical trials reduce the likelihood that there may be a systematic difference between the treatment and control groups (who may be given placebos). A well designed study should be reproducible by other researchers. Unfortunately lack of reproducibility can be a problem as reported on the [Science Exchange blog]: "Amgen found that 47 of 53 “landmark” oncology publications could not be reproduced. Bayer found that 43 of 67 oncology & cardiovascular projects were based on contradictory results from academic publications. Dr. John Ioannidis and his colleagues found that of 432 publications purporting sex differences in hypertension, multiple sclerosis, or lung cancer, only one data set was reproducible." (b) Confounding. One thing to look out for is when two effects both occur at the same time so that you cannot know which was the real cause. This situation is called confounding. For example, since the main source of lycopenes is tomato products if one found that consumption of tomato products reduced the risk of prostate cancer one could never be quite sure if the reason was due to lycopenes or due to something else in the tomato or even due to something else that is typically eaten with tomato. For example, tomato is a common component of pizza and there is some test tube evidence that the carvacrol component of oregano may have an anti-cancer effect. The anti-cancer effect being hypothesized for lycopenes may actually be due to oregano (See [Science Daily]. We saw that lycopenes and other tomato components are confounded with each other. If the confounders are known and differentially expressed then it may be possible to eliminate their effect using statistical models. Allocating subjects randomly to the treatment or control groups has the advantage that confounders may be removed even when they are unknown. (c) Comparability of Treatment and Control Groups. The treatment and control groups should be composed of similar patients. Problems occur when the treatment group is healthier than the control group. In that case it might be that better performance with a drug on the treatment group is really just a measure of their superior health status and not related to the drug at all. In this case health status was a confounder since both heath status or treatment could explain any difference invalidating any conclusion that the result is due to the treatment. If the treatment and control groups are not comparable its important to adjust for differences. In prostate cancer studies stratifying or adjusting for differences in Gleason score or similar categorizations of disease severity is typically an important consideration. (See [Pubmed: 12693887] [Full Free Text] for a for a non-technical primer on multivariable adjustment and stratification.) For example, let us again consider consider the May 2008 study [PMID: 18467718] that aims to compare laparascopic/robotic and open prostatectomy outcomes that was mentioned previously. The investigators "assessed the association between surgical approach and outcomes, adjusting for surgeon volume, age, race, comorbidity, and geographic region." However, as pointed out in [PMID: 18467714] [Full Text] the study also did not look at disease severity and "without clinical and pathologic data, one cannot reliably account for these differences". The same point is made in a May 10, 2008 Washington Post article [link] which says that the "study did not look at staging and scoring of the tumor, meaning that some of the differences seen might be due to differences in disease" rather than differences between laparoscopy and open surgery. Another example where adjustment was crucial was a study on health care workers who were accidentally pricked with HIV needles. Those treated with AZT (a medicine for HIV) fared no better than those who were not so treated; however, when adjustment was made for the severity of the needle prick then AZT administration was observed to have a beneficial effect. Had this adjustment not been made a potentially life saving treatment might have been overlooked. See the Neuroskeptic Blog and [PMID: 9366579] [Full Free Text]. Yet another comparability problem to be aware of are situations where the control treatment has a beneficial or detrimental effect. For example, researchers have hypothesized that the control treatment in the Provenge trials actually harmed the patients which would make the Provenge arm of the trial seem better than it really was. See [News].
- Analogy. If a certain phenomenon is found to exist in breast cancer, also a hormone based cancer, it would make its appearance in prostate more plausible. If one COX-2 inhibitor has a certain side effect then it is more plausible that another one will too.
. A presentation on conflict of interest in the pharmaceutical industry (funded by a grant from the Oregon state Attorney General Consumer and Prescriber Education Program which is funded by the multi-state settlement of consumer fraud claims regarding the marketing of the prescription drug Neurontin) can be found [here]. Among the many point it brings out are:
- Spingarn et al [PMID: 8540971], conducted a retrospective cohort study of hospital resdients who attended an industry sponsored grand rounds and their views related to treatment options for Lyme Disease. 3 months after, residents who attended were more likely to choose the manufacturer’s product and less likely to select scientifically preferred antibiotic over sponsored product.
- Bekelman et al [PMID: 12533125] performed a meta analysis which found that industry sponsored studies were 3.6x more likely to favor industry.
- Nieto et al [PMID: 17954797] compared the Adverse effects of inhaled corticosteroids in funded and nonfunded studies and found that industry sponsored studies were more likely to employ design features that were less likely to detect differences in adverse effects and that industry funded trials were 4x more likely to conclude drug is safe given statistically significant differences in adverse effects.
Bradford Hill in PracticeLooking at how these are used, in this 1999 paper [PMID: 10359231] [Full Text] the authors write: "In practice, the criteria of consistency, strength of association, dose response, and plausibility are used frequently and in that order; whereas, in the methodologic literature the criteria of strength of association, temporality, consistency, plausibility, dose response, and specificity are most often mentioned (in descending order)."
Readers may also wish to look at the original presentation by Hill and various discussions of it in the References section below.
Counterfactual CausationHill's causation is sometimes referred to as counterfactual causation. That means a cause which makes a difference in the effect when it is present (referred to as the factual case) versus when it is absent (referred to as the counterfactual case or potential outcome -- i.e. if a patient has prostate cancer then the counterfactual case would be the situation where that same patient does not have prostate cancer -- the counterfactual case can never be observed because any one patient either has prostate cancer or does not not). This causation is generally interpreted in a statistical sense -- that is, the cause increases the probability of the effect but is neither necessary nor sufficient. Not all smokers get lung cancer yet those who do not smoke have lower rates of the disease than those who do. This is further discussed by Page et al in the references below.
Example - Vitamin DIn a 2008 study of Finnish smokers Ahn et al (2008) [PMID: 18505967] [Full text] found that the 20% of patients with the lowest blood levels of vitamin D has also had the lowest risk of prostate cancer. There are several reasons to be suspicious of this. Firstly, the dose response criterion is not satisfied. That prostate cancer risk did not always increase with vitamin D level as one proceeds from one quintile to the next. (The quintiles refer to the 5 groups such that the first quintile is the 20% of patients with the lowest vitamin D in their blood, the second quintile is the 20% with the next lowest, etc.) There is also a worry here about temporality and the associated reverse causation. In particular, perhaps those who had the highest vitamin D levels were sicker and were trying actively to prevent further disease through intake of vitamin D?
Example - Fish IntakeWe apply the Bradford Hill criteria to the effect of fish intake on survival among those diagnosed with prostate cancer.
In a 22 year prospective cohort study among 20,167 men participating in the Physician's Health Study who were free of cancer in 1983 found that 2161 men were later diagnosed with prostate cancer and 230 died of prostate cancer. A November 2008 study published in the American Journal of Clinical Nutrition [PMID: 18996866] [Partial Text] has concluded that fish "intake is unrelated to prostate cancer incidence" (consistent with prior studies); however, they also found that "Higher intakes of fish, particularly dark meat fish, and seafood n-3 fatty acids were related to lower prostate cancer mortality among the men diagnosed with prostate cancer." We examine this latter finding in terms of the Bradford Hill criteria:
Strength. First note that the finding was quite strong as "survival analysis among the men diagnosed with prostate cancer revealed that those consuming fish greater than or equal to 5 times/wk had a 48% lower risk of prostate cancer death than did men consuming fish less than once weekly [relative risk (RR) = 0.52; 95% CI: 0.30, 0.91; P for trend = 0.05]."
Consistency. The study appears to be consistent with related results. From the paper: "In a recent small pilot randomized trial, men assigned to a study diet that emphasized, among other changes, increased intake of n-3 fatty acid-rich fish, had an increase in PSA doubling time [PMID: 9182974] [Full Text]. Similarly, the studies that previously reported fish or long-chain n-3 fatty acid intake to decrease prostate cancer risk, including our previous work [PMID: 17585059], have generally reported stronger associations with advanced stage [PMID: 10584888], [PMID: 15213050] [Full Text], [PMID: 15495177], clinically aggressive [PMID: 17585059], metastatic [PMID: 12540506] [Full Text], or lethal [PMID: 11403817] disease, which suggests that these dietary factors may have a role in reducing disease progression or mortality.
Temporality. The diet in the study was measured at baseline -- that is, at the beginning of the study when the patients did not have prostate cancer.
Biological Plausibility. The paper points out the biological plausibility as follows: "Laboratory data also suggest a role of marine n-3 fatty acids in reducing prostate cancer progression and mortality. Eicosapentaenoic acid, and to a greater extent its 15-LOX metabolite 15-HEPE, suppress the proliferation of multiple prostate cancer lines and the generation of COX-2 and 5-LOX metabolites of arachidonic acid [PMID: 15850718] that are known to increase proliferation, tumor cell survival, and angiogenesis[PMID: 15159222] [Full Text, [PMID: 16289380], [PMID: 9789062], [Full Text] , [PMID: 9199209]. Moreover, in a mouse model simulating prostate cancer recurrence after radical prostatectomy, mice fed an eicosapentaenoic acid precursor had reduced tumor recurrence, increased PSA doubling time, and decreased proliferation and increased apoptosis in recurrent tumor cells (14) Likewise, mice fed eicosapentaenoic acid in a model of hormone ablation therapy showed improved response to therapy (higher tumor apoptosis-to-mitosis ratios) and decreased progression into androgen independence (45)."
Coherence. The fact that fish intake seems unrelated to prostate cancer risk yet related to progression might seem contradictory yet observations that prostate cancer takes on different characteristics at different stages of its progression mean that this is certainly a possibility. See Willett Divides Prostate Cancer into Four. On the other hand, correlations between omega-3 fatty acids in the blood and fish intake were high.
Specificity. The study did attempt to control for other possible influences of prostate cancer mortality including "age at prostate cancer diagnosis; BMI; physical activity; intakes of alcohol, tomato products, and dairy products; smoking; race; use of multivitamins; use of vitamin E supplements; random assignment to aspirin or ß-carotene; and tumor stage and grade at diagnosis". It was unable to adjust for energy (calorie) intake but since it did adjust for BMI this can be considered a reasonable surrogate."
Experimental Evidence. This was not a randomized controlled study so the evidence must be considered as suggestive only. (Actually the data is from a randomized trial for asprin and beta carotene but not for this purpose so relative to this purpose it was not randomized.)
The strength of association, consistency with other studies and biological plausibility suggest mean that the study in question is highly suggestive of reduced fat intake having a desirable effect on prostate cancer survival among those diagnosed with prostate cancer.
Example - LycopenesA 2002 review of evidence by Miller et al regarding lycopene and prostate cancer that made specific use of the Bradford Hill criteria appeared in [PMID: 12424328] [Full Text] and we shall summarize its findings.
Regarding strength of association, a reduction in risk by about 20% to 40% has been found in larger well-controlled lycopene and prostate cancer studies. 8 of 10 reviewed studies including 47,000 person Health Professional's Follow-Up Study (HPFS) found a relationship suggesting consistency in findings. Dose response was shown in the HPFS study. Cancer develops over many years and it is generally not feasible to trace the entire lifetime eating habits of subjects so whether an observed diet high in lycopene preceded the disease in the case of prostate cancer is certainly open to question. Specificity is also questionable since there may be many causes for prostate cancer. Also it is difficult to separate the effect of lycopenes from tomato in general since tomato products are the primary source of Lycopenes in the American diet and an effect of lycopenes could be due to something else in tomato products. Biological mechanisms have been found that suggest the plausibility of an anti-cancer effect:
the 11 conjugated and two non-conjugated double bonds in lycopene make it highly reactive towards oxygen and free radicals, and this anti-oxidant activity probably contributes to its efficacy as a chemoprevention agent. The reactivity of lycopene also explains why it isomerizes rapidly in blood and tissues from the biosynthetic all-trans form to a mixture of cis-isomers. ... In addition to antioxidant activity, in vitro experiments indicate other mechanisms of chemoprevention by lycopene including induction of apoptosis and antiproliferation in cancer cells, anti-metastatic activity, and the upregulation of the antioxidant response element leading to the synthesis of cytoprotective enzymes. [PMID: 18585855]Human trials have shown mixed results.
Overall there is substantial evidence in favor of a protective effect of lycopenes yet it is not entirely conclusive. The Miller et al paper points out that "scientists, regulatory agencies, marketers of products, and those defining public health policy have differing opinions regarding the strength of the data when applied to criteria for inference and causality." In a 2005 review, Kristal and Schenk [PMID: 16046734] "do not find evidence to support a substantial association of lycopene or tomato products with reduced prostate cancer risk. Results of large cohort studies are mixed, and few serum-based or case-control studies were sufficiently large to detect modest effect sizes. Most of the studies published on lycopene and cancer have not been designed or analyzed to answer this question optimally, and key areas for advancing this research include a) improving assessments of lycopene exposure; b) incorporating information on cancer stage, grade, and screening into statistical analyses; and c) investigating interactions between oxidative stress or polymorphisms in genes affecting response to oxidative stress with lycopene intake." In a May 2008 news item [link] Kristal was further quoted as saying: "The excitement about lycopene (the pigment that puts the red in tomatoes) and prostate-cancer prevention is probably a mistake ... Most large and well-designed human studies, and most animal studies, have failed to find convincing associations between lycopene and reduced prostate-cancer risk." A 2009 review by researchers also assessed the evidence as inconclusive. [Abstract] [PMID: 19901932]. A 2014 Cochrane review "identified 3 relevant studies, comprising 154 participants in total. Two of the studies were assessed to be of 'high' risk of bias. Meta-analysis of two studies indicated no statistical difference in prostate specific antigen (PSA) levels between men randomised to receive lycopene and the comparison group (MD -0.34, 95% CI -2.01 to 1.32). None of the studies assessed prostate cancer mortality. No other meta-analyses were possible since other outcomes assessed only had one study contributing data."
One potential explanation is that the way lycopene is combined with other ingredients is key to whether or not it is effective. A news item reporting an upcoming study to be published in Cancer Research found that working with certain rats that 10% of rats fed dehydrated tomato paste had prostate tumors vs. 30% for regular tomato paste and 60% with no tomato paste. Human trials would be required to determine if this result continues to hold. The researchers believe that FruHis in dehydrated tomato paste may be a key element. See [link]. In animal studies and small short term human studies Fleshner found improved response from combining lycopene with other antioxidants relative to lycopene itself. See this post. Thus it may be that lycopene must be administered in a certain way or in conjunction with certain other ingredients in order to find a beneficial effect.
For more detailed information on lycopenes see the US FDA site and their explanation of why they found the evidence insufficient to allow health claims for lycopenes: [link]
Historical Note. Sir Austin Bradford Hill is regarded as the inventor of randomized clinical trials. He was influenced by Sir R. A. Fisher and David Hume who pioneered the idea of randomization. [link]
ReferencesSir Austin Bradford Hill, The environment and disease: association or causation?, Proceedings of the Royal Medical Society of Medicine, May 58 (1965), 295-300. [PMID: 14283879] [Full Text] [Full Text]
Epidemiology, Wikipendia. See section entitled Bradford-Hill criteria. [link]
Hofer, A. [PMID: 16269083][Full Text]
Phillips and Goodman [link] write that Hill intended that his criteria be a set of guidelines rather than a formal set of rules.
Page et al has a discussion causation in a medical context [link]