It is a selection bias. Suppose 1000 people are going through a psychotic crisis in families refusing psychiatric treatment. Suppose, 5 years later, 90% have recovered, and 10% have worsened. In desperation, the families of these 10% finally hospitalize them. In this case, we observe that 100% of people hospitalized have worsened in the last 5 years. Psychiatrists might observe that, in this group of chronic psychotics, the recovery rate is only 5%. But it is a selection bias: the original sample of this group was 1,000 people, and 90% of this group have recovered and will never go to psychiatry. They are therefore invisible. Now suppose 100 people are going through a psychotic crisis, but this time, in families following the recommendations of psychiatrists, and immediately hospitalizing their loved one. Suppose that with medication, 5 years later, 30% of people recover and 70% become chronic psychotics. Thus, according to psychiatric observations: 30% of psychotics treated immediately recover, 40% become chronic; 5% of psychotics treated 5 years later recover, 95% remain chronic. But according to the actual data: 90% of psychotics never treated recover, 10% become chronic; 30% of psychotics treated immediately recover, 70% become chronic. – Association ≠ causality. Sometimes a negative association can reveal positive causation.